BEGIN:VCALENDAR CALNAME:AI Engineer World's Fair 2025 NAME:AI Engineer World's Fair 2025 PRODID:-//github.com/ical-org/ical.net//NONSGML ical.net 4.0//EN VERSION:2.0 X-WR-CALNAME:AI Engineer World's Fair 2025 BEGIN:VTIMEZONE TZID:Pacific Standard Time X-LIC-LOCATION:America/Los_Angeles BEGIN:STANDARD DTSTART:20241103T020000 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:PST TZOFFSETFROM:-0700 TZOFFSETTO:-0800 END:STANDARD BEGIN:DAYLIGHT DTSTART:20250309T020000 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:PDT TZOFFSETFROM:-0800 TZOFFSETTO:-0700 END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT DESCRIPTION: DTEND:20250603T094500 DTSTAMP:20260403T174839Z DTSTART:20250603T071500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Continental Breakfast UID:SZSESSION914b9c82-2913-4ded-8c02-97f30c758261 END:VEVENT BEGIN:VEVENT DESCRIPTION:Registration is open from 8:00am-7:00pm in the Event Hub DTEND:20250603T190000 DTSTAMP:20260403T174839Z DTSTART:20250603T080000 LOCATION:Atrium: Event Hub SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Registration UID:SZSESSIONa32dc29f-15e0-4431-819f-78037834eebf END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Comfy Anonymous\n\nQuick introduction to ComfyUI and what's new followed by a QA session. DTEND:20250603T095000 DTSTAMP:20260403T174839Z DTSTART:20250603T090000 LOCATION:Salons 2-6: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:ComfyUI UID:SZSESSION947560 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ishan Anand\n\nDon't be intimidated. Modern AI can fe el like magic\, but underneath the hood are principles that web developer s can understand\, even if you don't have a machine learning background. In this workshop\, we'll explore a complete GPT-2 inference implementatio n built entirely in Vanilla JS. This JavaScript translation of the popula r "Spreadsheets-are-all-you-need" approach will let you debug and step th rough a real LLM line by line without the overhead of learning a new lang uage\, framework\, or even IDE.\n\nAll the major LLMs\, including ChatGPT \, Claude\, DeepSeek\, and Llama\, inherit from GPT-2's architecture\, ma king this exploration a solid foundation to understand modern AI systems and comprehend the latest research.\n\nWhile we won't have time to cover *everything*\, you'll gain the essential knowledge to understand the key concepts that matter when building with LLMs\, including how they:\n\n-Co nvert raw text into meaningful tokens\n- Represent semantic meaning throu gh vector embeddings\n- Train neural networks through gradient descent\n- Generate text with sampling algorithms like top-k\, top-p\, and temperat ure\n\nThis intense but beginner-friendly workshop is designed specifical ly for web developers diving into ML and AI for the first time. It’s your "missing AI degree" in just two hours. You'll walk away with an intuitiv e mental model of how Transformers work that you can apply immediately to your own LLM-powered projects. DTEND:20250603T110000 DTSTAMP:20260403T174839Z DTSTART:20250603T090000 LOCATION:Golden Gate Ballroom A: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How LLMs work for Web Devs: GPT in 600 lines of Vanilla JS UID:SZSESSION915928 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Nina Lopatina\, Rajiv Shah\n\nWant to take advantage of your data\, but don't want to reinvent RAG infrastructure? Join our w orkshop and see how you can deploy Agentic RAG in minutes using Contextua l AI's managed RAG solution. We'll explore how Contextual handles intelli gent parsing and chunking of your data\, retrieves information with state of the art accuracy\, and generates responses with a multi layered set o f guardrails against hallucinations. Together\, we'll build an end-to-end Agentic RAG pipeline and demonstrate its integration with Claude Desktop via MCP\, so you can see how this could plug into your existing ecosyste m. \n\nBy the end of this session\, you'll have a functioning Agentic RAG prototype that you can easily customize and deploy to production for you r specific use cases\, even with complex\, unstructured documents. DTEND:20250603T102000 DTSTAMP:20260403T174839Z DTSTART:20250603T090000 LOCATION:Golden Gate Ballroom B: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Forget RAG Pipelines—Build Production-Ready AI Agents in 15 Minutes UID:SZSESSION936933 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Zach Blumenfeld\n\nLearn the foundations of GraphRAG\ , starting with knowledge graph construction and then common retrieval pa tterns. DTEND:20250603T102000 DTSTAMP:20260403T174839Z DTSTART:20250603T090000 LOCATION:Golden Gate Ballroom C: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Intro to GraphRAG UID:SZSESSION933714 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Daniel Han\n\nWhy is Reinforcement Learning (RL) sudd enly everywhere\, and is it truly effective? Have LLMs hit a plateau in t erms of intelligence and capabilities\, or is RL the breakthrough they ne ed?\n\nIn this workshop\, we'll dive into the fundamentals of RL\, what m akes a good reward function\, and how RL can help create agents.\n\nWe'll also talk about kernels\, are they still worth your time and what you sh ould focus on. And finally\, we’ll explore how LLMs like DeepSeek-R1 can be quantized down to 1.58-bits and still perform well\, along with techni ques to maintain accuracy. DTEND:20250603T120000 DTSTAMP:20260403T174839Z DTSTART:20250603T090000 LOCATION:Foothill C: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Advanced: Reinforcement Learning\, Kernels\, Reasoning\, Quantizat ion & Agents UID:SZSESSION929509 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Philip Kiely\, Yineng Zhang\n\nDo you want to learn how to serve models like DeepSeek and Qwen with SOTA speeds on launch day ? SGLang is an open-source fast serving framework for LLMs and VLMs that generates trillions of tokens per day at companies like xAI\, AMD\, and M eituan. This workshop guides AI engineers who are familiar with serving m odels using frameworks like vLLM\, Ollama\, and TensorRT-LLM through depl oying and optimizing their first model with SGLang\, as well as providing guidance on when SGLang is the appropriate tool for LLM workloads. DTEND:20250603T102000 DTSTAMP:20260403T174839Z DTSTART:20250603T090000 LOCATION:SOMA: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Introduction to LLM serving with SGLang UID:SZSESSION916066 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Damien Murphy\n\nEver wished your webhooks could thin k for themselves? Join us to discover how A2A agents can transform passiv e webhook endpoints into intelligent workflow processors.\n\nIn this sess ion\, we'll show you how to build a system that automatically spawns AI A gents to handle incoming webhooks. \n\nUsing Google's Agent-to-Agent fram ework and MCP\, you'll learn how to create dynamic AI agents that respond to events\, communicate with external services\, and make decisions base d on content analysis.\n\nSee the future of workflow automation where web hooks don't just trigger actions—they trigger intelligence! DTEND:20250603T102000 DTSTAMP:20260403T174839Z DTSTART:20250603T090000 LOCATION:Foothill G1&2: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:A2A & MCP: Automating Business Processes with LLMs UID:SZSESSION911821 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Taylor Jordan Smith\n\nAccuracy scores and leaderboar d metrics look impressive—but production-grade AI requires evals that ref lect real-world performance\, reliability\, and user happiness. Tradition al benchmarks rarely help you understand how your LLM will perform when e mbedded in complex workflows or agentic systems. How can you realisticall y and adequately measure reasoning quality\, agent consistency\, MCP inte gration\, and user-focused outcomes?\n\nIn this practical\, example-drive n talk\, we'll go beyond standard benchmarks and dive into tangible evalu ation strategies using various open-source frameworks like GuideLLM and l m-eval-harness. You'll see concrete examples of how to create custom eval suites tailored to your use case\, integrate human-in-the-loop feedback effectively\, and implement agent reliability checks that reflect product ion conditions. Walk away with actionable insights and best practices for evaluating and improving your LLMs\, ensuring they meet real-world expec tations—not just leaderboard positions! DTEND:20250603T102000 DTSTAMP:20260403T174839Z DTSTART:20250603T090000 LOCATION:Nobhill A&B: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Beyond Benchmarks: Strategies for Evaluating LLMs in Production UID:SZSESSION915684 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Chelcie Taylor\, Dani Grant\n\nA no fluff\, all tact ics discussion. More AI engineers should build startups\, the world needs more software. But there’s a way to raise VC and it’s hard to do it if y ou’ve never seen it done. We are going to walk through the exact playbook to raise your first round of funding. We will show you real pitch decks\ , real cold emails and real term sheets so when you go out to raise your first round of funding\, you are setup to do it. Every AI Engineer should be equip to start their own company and this session makes sure raising $$$ is not going to be the blocker. DTEND:20250603T102000 DTSTAMP:20260403T174839Z DTSTART:20250603T095500 LOCATION:Salons 2-6: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The AI Engineer’s Guide to Raising VC UID:SZSESSION913965 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250603T170000 DTSTAMP:20260403T174839Z DTSTART:20250603T100000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Rehearsals/Tech Check UID:SZSESSIONabd66601-8f27-42ce-8a87-1d246de8a58c END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250603T104500 DTSTAMP:20260403T174839Z DTSTART:20250603T101500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Morning Break UID:SZSESSIONa03323dd-5567-48c3-824a-eed2c5266a09 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Kwindla Kramer\, Shrestha Basu Mallick\n\nVoice AI A gents are being deployed today in a wide range of business contexts. For example:\n\n - handling an increasing variety of call center tasks\,\n - collecting patient data prior to healthcare appointments\,\n - followi ng up on inbound sales leads\,\n - coordinating scheduling and logistics between companies\, and\n - answering the phone for nearly every kind o f small business.\n\nOn the consumer side\, conversational voice (and vid eo) AI is also starting to make its way into social applications and game s. And developers are sharing innovative personal voice AI projects and e xperiments every day on GitHub and social media.\n\nBuilding production-r eady voice agents is complicated. Many elements are non-trivial to implem ent from scratch.\n\nThis workshop will start with an overview of the voi ce AI landscape today.\n\n - The models\, APIs\, and infrastructure are used for Voice AI applications that are operating at production scale. \n - How to write voice agent code that achieves ultra low latency convers ation and enterprise-quality reliability.\n - What new models and tools are coming in the second half of 2025.\n\nThen we will shift to a hands-o n format: build and deploy a voice agent.\n\nEngineers from Google and Da ily will help you get set up with a starter kit repo for your intended us e case\, then help you extend that code to create your own\, customized V oice AI application. DTEND:20250603T120000 DTSTAMP:20260403T174839Z DTSTART:20250603T104000 LOCATION:Salons 2-6: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Voice Agents with Gemini and Pipecat UID:SZSESSION916131 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Corey Cooper\n\nThis workshop explores how USDC\, AI\ , and smart contracts can streamline escrow by automating fund release ba sed on task or process verification. By using AI to interpret off-chain s ignals such as document validation\, delivery confirmations\, or mileston e completion\, we can trigger secure\, programmable USDC payouts without manual intervention. The result is a faster\, trust-minimized escrow syst em ideal for services\, trade\, and gig economy use cases. DTEND:20250603T120000 DTSTAMP:20260403T174839Z DTSTART:20250603T104000 LOCATION:Golden Gate Ballroom B: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Automating Escrow with USDC and AI UID:SZSESSION936937 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Doug Guthrie\n\nThis hands-on workshop will guide par ticipants through the complete AI evaluation lifecycle using Braintrust\, from initial prompt testing to production monitoring. Attendees will lea rn to build evaluation frameworks that ensure their AI applications perfo rm reliably in real-world scenarios. Topics covered include both offline and online evaluation strategies\, logging and feedback systems\, and hum an review processes. DTEND:20250603T120000 DTSTAMP:20260403T174839Z DTSTART:20250603T104000 LOCATION:Golden Gate Ballroom C: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Mastering AI Evaluation: From Playground to Production with Braint rust UID:SZSESSION942803 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Apoorva Joshi\n\nIn this hands-on workshop\, you will build a multimodal AI agent capable of processing mixed-media content—fr om analyzing charts and diagrams to extracting insights from documents wi th embedded visuals. Using MongoDB as a vector database and memory store\ , and Google's Gemini for multimodal reasoning\, you will gain hands-on e xperience with multimodal data processing pipelines and agent orchestrati on patterns by implementing core components directly\, using good ol' Pyt hon.\n DTEND:20250603T120000 DTSTAMP:20260403T174839Z DTSTART:20250603T104000 LOCATION:SOMA: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Multimodal AI Agents (From Scratch) UID:SZSESSION933707 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: David Karam\n\nOne of the biggest challenges in build ing evals you can trust is building metrics that reliably measure goodnes s in your application\; metrics that are highly accurate\, rapid fast\, a nd tunable to ground truth rater and user behavior. This workshop is insp ired by decades of AI and machine learning development in Google Search\, reinvented for the modern LLM stack by the Pi team over the past year.\n \nIn this workshop you will learn how to:\n1) Brainstorm and design custo m metrics tailored to your specific application needs.\n2) Identify which types of signals (natural language\, code\, other models) work best for your use case through rapid trial and error.\n3) Combine & calibrate your metrics against ground truth data using real examples from your domain.\ n4) Use simple tools like Google Sheets for visualizing and analyzing you r inputs and outputs with those metrics.\n5) Integrate your scoring model s into both online workflows like agent control and offline ones like mod el comparison and training evaluation.\n\n DTEND:20250603T120000 DTSTAMP:20260403T174839Z DTSTART:20250603T104000 LOCATION:Foothill G1&2: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Solving for the hardest Eval challenge: Building Metrics that actu ally work UID:SZSESSION907684 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Jeremy Adams - Casañas\, Kyle Penfound\n\nCoding age nts are transforming how software gets built\, tested\, and deployed\, bu t engineering teams face a critical challenge: how to embrace this automa tion wave without sacrificing trust\, control\, or reliability.\nIn this 80 minute workshop\, you’ll go beyond toy demos and build production-mind ed AI agents using Dagger\, the programmable delivery engine designed for real CI/CD and AI-native workflows. Whether you're debugging failures\, triaging pull requests\, generating tests\, or shipping features\, you'll learn how to orchestrate autonomous agents that live in and around your codebase: from your laptop to your CI platform.\nWe’ll guide you through: \n\nBuilding real-world agents with Dagger and popular LLMs (GPT\, Claude \, etc.)\n\nProgramming agent environments using real languages (Go\, Pyt hon\, TypeScript)\n\nExecuting agent workflows locally and in GitHub Acti ons\, so you can bring them to production\n\nUsing a composable runtime t hat ensures isolation\, determinism\, traceability\, and repeatability\n\ nDesigning agents that automate and enhance debugging\, test generation\, code review\, bug fixing\, and feature implementation\n\n\nBy the end of the workshop\, you’ll walk away ready to build your own army of autonomo us agents\, working collaboratively across your codebase\, locally and in CI\, accelerating development without ceding control. Let’s build agents that don’t just talk\, they ship! DTEND:20250603T120000 DTSTAMP:20260403T174839Z DTSTART:20250603T104000 LOCATION:Nobhill A&B: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Ship Agents that Ship: A Hands-On Workshop for SWE Agent Builders UID:SZSESSION915961 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Christopher Harrison\n\nThe agent capabilities added to GitHub Copilot have enhanced its ability to act as a peer programmer. Copilot can now discover and generate code based on existing standards\, run tests\, recover from errors\, and call tools using Model Context Prot ocol (MCP). This workshop will guide you through piloting Copilot's agent capabilities and how to best integrate with the most widely adopted AI c oding assistant in the world.\n\nKey takeaways include:\n\n- Understandin g how and when to bring agents into your software development workflow\n- Providing context through the use of custom instructions and prompt file s to ensure consistency across your team\n- Discovering how MCP provides access to an additional set of external tools and capabilities that the a gent can use\n- Configuring Copilot's agentic capabilities to take advant age of your custom MCP server\n- Recommended best practices to help your responsibly accelerate your development while maintaining code quality an d governance DTEND:20250603T120000 DTSTAMP:20260403T174839Z DTSTART:20250603T104000 LOCATION:Nobhill C&D: Microsoft SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Piloting agents in GitHub Copilot UID:SZSESSION935459 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Thor 雷神 Schaeff\n\nIn this workshop you will learn ho w to build multilingual Conversational AI agents that can automatically d etect your user's spoken language and can seamlessly switch to their pref erred language. DTEND:20250603T120000 DTSTAMP:20260403T174839Z DTSTART:20250603T110000 LOCATION:Golden Gate Ballroom A: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Build multilingual Conversational AI Agents UID:SZSESSION915431 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250603T130000 DTSTAMP:20260403T174839Z DTSTART:20250603T120000 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Lunch UID:SZSESSION31aee5ff-2c08-4063-9d8a-600eae415a30 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Danielle Perszyk\n\nWe’re all hearing that AI agents will enable AGI\, but they can’t yet reliably perform even basic computer tasks. It turns out that getting AI to click\, type\, and scroll is more challenging than getting it to generate code. How can we build general-p urpose agents that can do anything we can do on a computer? \n\nThis is o ur goal at the Amazon AGI SF Lab. In this talk\, I’ll propose a new appro ach to agents that we call Useful General Intelligence. After describing how we’re solving the biggest challenges in computer use while enabling d evelopers to access our tech in it’s earliest developmental stages\, I’ll show real workflows that developers have built with Nova Act\, our agent ic model and SDK. DTEND:20250603T130000 DTSTAMP:20260403T174839Z DTSTART:20250603T124000 LOCATION:Golden Gate Ballroom A: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Useful General Intelligence UID:SZSESSION929790 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Duan Lightfoot\n\nIn this 2-hour workshop\, participa nts will gain practical hands-on experience building sophisticated AI age nts using Amazon's agent technologies. You'll learn to build agents that can navigate the web like humans\, perform complex multi-step tasks\, and leverage specialized tools through natural language commands. You’ll exp lore Amazon Nova Act for reliable web navigation\, Model Context Protocol (MCP) for connecting agents to external data sources and APIs\, and Amaz on Bedrock Agents for orchestrating complex workflows. Through guided exe rcises\, you'll create agents capable of retrieving information and takin g action across web applications\, all through natural language interacti ons. By the end of this workshop\, you'll have the practical skills to bu ild AI agents that can browse websites\, interact with web interfaces\, a nd solve multi-step problems by combining these powerful Amazon technolog ies. DTEND:20250603T150000 DTSTAMP:20260403T174839Z DTSTART:20250603T130000 LOCATION:Golden Gate Ballroom A: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Agents with Amazon Nova Act and MCP UID:SZSESSION933607 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ilan Bigio\n\nCovering all forms of fine-tuning and p rompt engineering\, like SFT\, DPO\, RFT\, prompt engineering / optimizat ion\, and agent scaffolding. DTEND:20250603T150000 DTSTAMP:20260403T174839Z DTSTART:20250603T130000 LOCATION:Golden Gate Ballroom B: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Model-Maxxing: RFT\, DPO\, SFT (Fine-tuning with OpenAI) UID:SZSESSION930540 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Alison Cossette\, Andreas Kollegger\n\nAdvanced Grap hRAG techniques apply graph ML and algorithms\, wrapped into tidy noteboo ks. DTEND:20250603T150000 DTSTAMP:20260403T174839Z DTSTART:20250603T130000 LOCATION:Golden Gate Ballroom C: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Graph Intelligence: Enhance Reasoning and Retrieval Using Graph An alytics UID:SZSESSION921229 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Dan Mason\n\nWe've all seen website chat bots which c an look up an order or answer a basic question -- but what does it take t o build autonomous agents which manage long\, delicate processes like mul ti-day medical treatments? \n\nIn this workshop\, we'll explore a workfl ow Stride built in partnership with Avila (https://avilascience.com/) tha t helps patients self-administer medication regimens at home. The stack includes LangGraph/LangSmith\, Claude\, MCP\, Node.js\, React\, MongoDB\, and Twilio\, and rests on a foundation of treatment "blueprints" which L LM-powered agents use to guide patients to good outcomes. \n\nYou'll lea rn how to:\n-Build a hybrid system of code and prompts that leverages LLM decisioning to drive a web application\, message queue and database\n-De sign and maintain flexible agentic workflow blueprints\, with no special tools (just Google Docs!)\n-Create an agent evaluation system\, which use s LLM-as-a-judge to evaluate the complexity of each interaction and escal ate to human support when needed\n\nWe'll also talk about the prompt engi neered guidelines and guardrails which helps agents adhere to protocol as much as possible\, while gracefully handling curveballs from the patient . Please bring questions -- we look forward to sharing our learnings on how to make agentic systems like this work in the real world! DTEND:20250603T150000 DTSTAMP:20260403T174839Z DTSTART:20250603T130000 LOCATION:Foothill C: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Case Study + Deep Dive: Telemedicine Support Agents with LangGraph /MCP UID:SZSESSION907544 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Philipp Schmid\n\nHands on Workshop on learning to us e Gemini 2.5 Pro in combination with Agentic tooling and MCP Servers. DTEND:20250603T150000 DTSTAMP:20260403T174839Z DTSTART:20250603T130000 LOCATION:SOMA: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:AI Engineering with the Google Gemini 2.5 Model Family UID:SZSESSION910732 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Philipp Krenn\n\nVector search is only a feature. Sea rch engines and information retrieval have retaken their position as the foundation of RAG. This workshop takes you through decades of research\, what has been working for a long time\, and how it got better with Machin e Learning. DTEND:20250603T150000 DTSTAMP:20260403T174839Z DTSTART:20250603T130000 LOCATION:Foothill G1&2: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Information Retrieval from the Ground Up UID:SZSESSION916214 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Daniel Kim\, Daria Soboleva\n\nOur hands-on workshop will walk you through how to build your own Mixture of Agents (MoA) syst em using the fastest\, and most capable open models available: Qwen3-32B and Llama 3.3-70B. MoA is an emerging architecture that combines the stre ngths of multiple large language models in a layered\, agent-based design . This approach delivers superior performance by enabling specialized age nts to collaborate across layers—outperforming today’s frontier models in both accuracy and efficiency.\n\nTo ground this new paradigm in its root s\, we’ll also explore how Mixture of Experts (MoE) architectures continu e to push the boundaries of scale and specialization. Learn how Cerebras trains state-of-the-art MoEs from Daria Soboleva\, Head Research Scientis t. DTEND:20250603T150000 DTSTAMP:20260403T174839Z DTSTART:20250603T130000 LOCATION:Nobhill A&B: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:From Mixture of Experts to Mixture of Agents … with Super Fast Inf erence UID:SZSESSION939088 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Cedric Vidal\n\nThis workshop offers a hands-on intro duction to developing Large Language Model (LLM)-powered AI agents using Microsoft’s Azure AI Agent Service. Participants will build a conversatio nal agent capable of analyzing sales data\, generating visualizations\, a nd delivering actionable insights.\n \nThe session takes a code-first app roach using the Azure AI Foundry SDK for Python\, and demonstrates how to integrate core Azure services including Azure OpenAI\, Azure AI Search\, and Azure Storage.\n \nAttendees will explore key concepts such as funct ion calling\, document grounding\, and leveraging code interpreters to ge nerate diagrams. The workshop also covers how to connect agents to extern al data sources like SQL databases (e.g.\, SQLite)\, enabling access to l egacy relational systems.\n \nBy the end of the session\, participants wi ll have a solid foundation for building and deploying intelligent\, code- first AI agents with Azure AI Agent Service—ready to power real-world app lications. DTEND:20250603T150000 DTSTAMP:20260403T174839Z DTSTART:20250603T130000 LOCATION:Nobhill C&D: Microsoft SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Code-First AI Agents with Azure AI Agent Service: A Pract ical introduction UID:SZSESSION936905 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250603T153000 DTSTAMP:20260403T174839Z DTSTART:20250603T150000 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Afternoon Break UID:SZSESSION24df35c2-0bac-4167-8f45-4245337a1d78 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Nick Nisi\, Zack Proser\n\nThis hands-on workshop in troduces Mastra.ai\, a TypeScript framework that streamlines the developm ent of agentic AI systems compared to traditional approaches using LangCh ain and vector databases. Participants will learn to build structured AI workflows with composable tools and reliable control\, enabling them to c reate internal AI assistants that can handle requests like data cleaning\ , email drafting\, and document summarization with minimal code. The sess ion covers Mastra installation\, running a local MCP server\, defining to ols and agents in TypeScript\, using the Mastra playground\, and implemen ting practical examples such as RAG setups and tool-chaining agents—all d esigned to equip attendees with the skills to develop scalable AI-driven internal tools based on sound software engineering principles rather than just experimental prompts. DTEND:20250603T173000 DTSTAMP:20260403T174839Z DTSTART:20250603T153000 LOCATION:Salons 2-6: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:AI Pipelines and Agents in Pure TypeScript with Mastra.ai UID:SZSESSION933688 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Numair Baseer\n\nAgentic coding marks a new era in so ftware development\, where AI agents take on autonomous roles in coding t asks. The Windsurf IDE embodies this shift by integrating intelligent age nts like Cascade\, which maintain full codebase context to perform multi- file edits\, run terminal commands\, and suggest changes through tools li ke Supercomplete and Flows. In this session\, we will explore features th at allow developers to guide strategy while the AI handles execution\, en hancing productivity and enabling more creative\, high-level work. DTEND:20250603T173000 DTSTAMP:20260403T174839Z DTSTART:20250603T153000 LOCATION:Golden Gate Ballroom A: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Agentic Coding with Windsurf UID:SZSESSION933632 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Dominik Kundel\n\nWe'll walk through the differences between chained and speech-to-speech powered voice agents\, how to approa ch them\, best practices and transform a text-based agent into our first voice-enabled agent DTEND:20250603T173000 DTSTAMP:20260403T174839Z DTSTART:20250603T153000 LOCATION:Golden Gate Ballroom B: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building voice agents with OpenAI UID:SZSESSION914537 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Carlos Esteban\n\nJoin us for a hands-on workshop whe re you'll learn practical strategies to evaluate AI applications througho ut their lifecycle—from initial testing of prompts to ongoing monitoring in production. We’re excited to host Sarah Sachs\, AI Lead at Notion\, wh o will share insights into how Notion built their acclaimed Notion AI. DTEND:20250603T173000 DTSTAMP:20260403T174839Z DTSTART:20250603T153000 LOCATION:Golden Gate Ballroom C: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How to build world-class AI products (featuring Sarah Sachs\, AI l ead @ Notion) UID:SZSESSION942858 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Forrest Brazeal\, Matt Ball\n\nAttendees will learn to use an AI coding agent as a fast and intuitive part of navigating and working with complex\, production-grade legacy code bases. We will drop d irectly into the code–written in assembly–that landed the1969 Apollo 11 a stronauts on the moon and\, through a series of challenges\, locate parts of the code tied to key functionality. Using the agent to convert a key guidance computer algorithm into a more modern programming language\, att endees will then compete to see whose code has what it takes to land on t he moon. DTEND:20250603T173000 DTSTAMP:20260403T174839Z DTSTART:20250603T153000 LOCATION:Foothill C: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Navigating deep context in legacy code with Augment Agent UID:SZSESSION933560 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Sander Schulhoff\n\nLearn from the creator of Learn P rompting\, the internet's 1st Prompt Engineering guide (released 2 months before ChatGPT)\, and HackAPrompt\, the World's 1st AI Red Teaming compe tition.\n\nMy talk will cover topics ranging from the history of prompt e ngineering to the most advanced research-backed prompt engineering techni ques.\n\nI will also discuss the origins of prompt injection and AI red t eaming\, as well as the current state of industry and the need for agenti c red teaming.\n\nFinally\, we will have an interactive competition where you will be able to hone your prompt hacking skills and win prizes from swyx!\n\nhttps://www.hackaprompt.com DTEND:20250603T173000 DTSTAMP:20260403T174839Z DTSTART:20250603T153000 LOCATION:SOMA: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Prompt Engineering & AI Red Teaming UID:SZSESSION947798 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Suman Debnath\n\nIn this workshop we will explore the integration of Colpali\, a cutting-edge Vision based Retrieval Model\, w ith voice synthesis for next-generation RAG systems. We'll demonstrate ho w Colpali's ability to generate multi-vector embeddings directly from doc ument images bypasses traditional OCR and complex preprocessing\, while a dding voice output creates a more intuitive and accessible user experienc e. Attendees will see how this combination handles documents with mixed t extual and visual information\, leading to more efficient and accurate in formation retrieval with natural voice responses. DTEND:20250603T173000 DTSTAMP:20260403T174839Z DTSTART:20250603T153000 LOCATION:Foothill G1&2: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:VoiceVision RAG - Integrating Visual Document Intelligence with Vo ice Response UID:SZSESSION933721 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Aman Khan\n\nGenAI is reshaping the product landscape \, creating huge opportunities (along with new expectations) for product managers. Yet while prompt engineering and model tuning get the spotlight \, one critical skill can get overlooked: rigorous evaluation.\n\nThis ta lk will help PMs move beyond gut-feel “vibe checks” to adopt concrete\, r epeatable evaluation strategies for LLM-powered products. I'll break down essential eval methodologies\, from human feedback and code-based checks to cutting-edge LLM-based evaluations. Drawing on real-world examples\, I'll share a practical framework PMs can use to:\n\n-Confidently evaluate AI-driven features\n- Ground decisions in real\, repeatable data\n- Buil d trust and delight through consistent quality\n DTEND:20250603T173000 DTSTAMP:20260403T174839Z DTSTART:20250603T153000 LOCATION:Nobhill A&B: Workshops SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Shipping AI That Works: An Evaluation Framework for PMs UID:SZSESSION915269 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Christopher Harrison\, Jon Peck\n\nGitHub Copilot's agentic capabilities enhance its ability to act as a peer programmer. Fro m the IDE to the repository\, Copilot can generate code\, run tests\, and perform tasks like creating pull requests using Model Context Protocol ( MCP). This instructor-led lab will guide you through using agent capabili ties on both the client and the server: Key takeaways include:\nUnderstan ding how to bring agents into your software development workflow\nIdentif ying scenarios where agents can be most impactful\, as well as tips and t ricks to provide the right context to lead to success\nDiscovering how Mo del Context Protocol provides access to an additional set of external too ls and capabilities that the agent can use\nRecommended practices to acce lerate your development while maintaining code quality. DTEND:20250603T173000 DTSTAMP:20260403T174839Z DTSTART:20250603T153000 LOCATION:Nobhill C&D: Microsoft SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Collaborating with Agents in your Software Development Workflow UID:SZSESSION940839 END:VEVENT BEGIN:VEVENT DESCRIPTION:The welcome reception will take place in the Expo Hall and the Grand Assembly. DTEND:20250603T190000 DTSTAMP:20260403T174839Z DTSTART:20250603T160000 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Welcome Reception UID:SZSESSIONc003876a-cc78-4550-8756-171c33114c4c END:VEVENT BEGIN:VEVENT DESCRIPTION:Organize your own in the app/AIE slack\, or sign up at https:/ /www.ai.engineer/#events DTEND:20250603T210000 DTSTAMP:20260403T174839Z DTSTART:20250603T173000 LOCATION:Atrium: Event Hub SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Community Meetups (Jun 3) UID:SZSESSION7d943dff-60b2-441b-8d35-d87255668ef1 END:VEVENT BEGIN:VEVENT DESCRIPTION:Registration is open from 7:00am-7:00pm in the Event Hub. DTEND:20250604T190000 DTSTAMP:20260403T174839Z DTSTART:20250604T070000 LOCATION:Atrium: Event Hub SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Registration UID:SZSESSION8c615cdb-afbd-42f4-9fec-fe66ad1ba51d END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250604T095500 DTSTAMP:20260403T174839Z DTSTART:20250604T071500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Continental Breakfast UID:SZSESSION74334729-a62e-4c05-b018-7d73009eb8dd END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250604T084500 DTSTAMP:20260403T174839Z DTSTART:20250604T074500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Rehearsals/Tech Check UID:SZSESSIONac9e5af0-736c-4dfa-8f34-5a699fd3c7a6 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250604T090000 DTSTAMP:20260403T174839Z DTSTART:20250604T084500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Keynote Doors Open UID:SZSESSION6f7d845c-bbab-4bf2-807b-c26fd46ee86a END:VEVENT BEGIN:VEVENT DESCRIPTION:Emcee Laurie Voss and curator swyx kick off the conference. DTEND:20250604T091000 DTSTAMP:20260403T174839Z DTSTART:20250604T090000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Welcome to AI Engineer UID:SZSESSIONc9b2b2f3-813a-4077-8319-61cec394dc9d END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: swyx .\n\nWhether you call it a workflow or an agent\ , AI engineered applications are seeing user-input:LLM-call ratios go fro m 1:1 (ChatGPT) to 1:100 (Deep Research\, Codex) and even 0:n (Ambient/Pr oactive agents). How does AI Engineering change as you build increasingly AI intensive applications? DTEND:20250604T092000 DTSTAMP:20260403T174839Z DTSTART:20250604T091000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Designing AI-Intensive Applications UID:SZSESSION949016 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Asha Sharma\n\nAI builders no longer ask whether to u se agents—but how many and how fast. In this kickoff keynote\, Microsoft’ s Asha Sharma shows what happens when natural language creation meets an industrial grade backbone. Watch live demos—to see agents move from idea to production in real time. Walk out with the commands\, repos\, and ope n protocols to build your piece of the agentic web. DTEND:20250604T094000 DTSTAMP:20260403T174839Z DTSTART:20250604T092000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Spark to System: Building the Open Agentic Web UID:SZSESSION936800 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Sarah Guo\n\ntba DTEND:20250604T100000 DTSTAMP:20260403T174839Z DTSTART:20250604T094000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:State of Startups and AI 2025 UID:SZSESSION927324 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Simon Willison\n\nWhat's changed in the world of LLMs since the AIE World's Fair last year? A lot! \n\nI'll be taking full adv antage of my role as a fiercely independent researcher to review the past 12 months of advances in the field and catch everyone up on the latest m odels\, free from any influence of vendors or employers. DTEND:20250604T102000 DTSTAMP:20260403T174839Z DTSTART:20250604T100000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:2025 in LLMs so far UID:SZSESSION935987 END:VEVENT BEGIN:VEVENT DESCRIPTION:Laurie introduces each track\, with special feature from Steph en Chin DTEND:20250604T103000 DTSTAMP:20260403T174839Z DTSTART:20250604T102000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Track Intros ft. Agentic GraphRAG UID:SZSESSIONc199b5b3-7635-4385-ac45-5ef9e3bb92c3 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250604T111500 DTSTAMP:20260403T174839Z DTSTART:20250604T103000 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Morning Break UID:SZSESSION1cda9955-fff1-4a76-8275-077cf1707d91 END:VEVENT BEGIN:VEVENT DESCRIPTION:Attendee-Only and Attendee-Led 10min lightning talks: see http s://crowdcomms.com/aiengineer25/qanda/41445 DTEND:20250604T105500 DTSTAMP:20260403T174839Z DTSTART:20250604T104000 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:[Hallway Track] Vision AI in 2025 — Peter Robicheaux\, Roboflow UID:SZSESSION08856acc-cac4-468d-806d-be5773c8a878 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Tomas Reimers\n\nThis talk will explore insights from millions of automated code reviews\, revealing trends in bugs\, vulnerab ilities\, and code health that Graphite’s AI code review agent have uncov ered. This talk will also provide meta commentary into the types of bugs AI code review agents are great at spotting\, and how far the field of AI code review has come in the last year alone. DTEND:20250604T105500 DTSTAMP:20260403T174839Z DTSTART:20250604T104000 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:AI-powered entomology: Lessons from millions of AI code reviews UID:SZSESSION933472 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Richmond Alake\n\nIn the rapidly evolving landscape o f agentic systems\, memory management has emerged as a key pillar for bui lding intelligent\, context-aware AI Agents. Inspired by the complexity o f human memory systems—such as episodic\, working\, semantic\, and proced ural memory—this talk unpacks how AI agents can achieve believability\, r eliability\, and capability by retaining and reasoning over past experien ces.\nWe’ll begin by establishing a conceptual framework based on real-wo rld implementations from memory management libraries and system architect ures:\nMemory Components representing various structured memory types (e. g.\, conversation\, workflow\, episodic\, persona)\nMemory Modes reflecti ng operational strategies for short-term\, long-term\, and dynamic memory handling\nNext\, the talk transitions to practical implementation patter ns critical for effective memory lifecycle management:\nMaintaining rich conversation history and contextual awareness\nPersistence strategies lev eraging vector databases and hybrid search\nMemory augmentation using emb eddings\, relevance scoring\, and semantic retrieval\nProduction-ready pr actices for scaling memory in multi-agent ecosystems\nWe’ll also examine advanced memory strategies within agentic systems:\nMemory cascading and selective deletion\nIntegration of tool use and persona memory\nOptimizin g performance around memory retrieval and LLM context window limits\nWhet her you're developing autonomous agents\, chatbots\, or complex workflow orchestration systems\, this talk offers knowledge and tactical insights for building AI that can remember\, adapt\, and improve over time.\nThis session is ideal for:\nAI engineers and agent framework developers\nArchi tects designing Agentic RAG or multi-agent systems\nPractitioners buildin g contextual\, personalized AI experiences\nBy the end of the session\, y ou’ll understand how to leverage memory as a strategic asset in agentic d esign—and walk away ready to build agents that not only act and reason bu t also remember. DTEND:20250604T105500 DTSTAMP:20260403T174839Z DTSTART:20250604T104000 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Architecting Agent Memory: Principles\, Patterns\, and Best Practi ces UID:SZSESSION933692 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Brian Johnson\, Chad Bailey\n\nTavus shipped the wor ld's first realtime video avatar platform last year. Developers use Tavus ' conversational video APIs to create education\, social\, and customer s upport agents. The Tavus team built their innovative product using the Pi pecat open source framework and Daily's global WebRTC infrastructure. Joi n us for a technical deep dive into conversational video. DTEND:20250604T105500 DTSTAMP:20260403T174839Z DTSTART:20250604T104000 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Realtime conversational video with Pipecat and Tavus UID:SZSESSION933493 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Tomas Reimers\n\nToday’s codebases move faster than e ver\, making human code reviews increasingly challenging. In this quick t echnical talk\, we’ll explore how AI-driven code review technology change s the game\, rapidly identifying subtle logic errors\, hidden security vu lnerabilities\, and performance bottlenecks before they ever reach produc tion. Drawing on insights from over millions of real-world pull requests reviewed by Diamond\, Graphite’s AI code review agent\, you'll learn how teams are achieving faster merges and higher code quality.\n\nWe'll wrap up with a brief introduction to Diamond\, demonstrating how effortlessly it integrates into your GitHub workflows\, followed by live demos and Q\& A with our engineers at the booth. DTEND:20250604T110000 DTSTAMP:20260403T174839Z DTSTART:20250604T104500 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Code Review for the Age of AI UID:SZSESSION933450 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Mark Myshatyn\n\nWhat does it mean to field not only LLMs\, but whole agentic solutions to highly regulated problems? Come joi n Los Alamos National Laboratory to hear about fielding AI in hard places . DTEND:20250604T111500 DTSTAMP:20260403T174839Z DTSTART:20250604T105500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Government Agents - AI Agents Meet Tough Regulations UID:SZSESSION915465 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Eashan Sinha\n\nAs experienced engineers\, especially senior and staff engineers\, our focus shifts towards complex problem-so lving\, architectural decisions\, and mentoring. While AI tools promise p roductivity gains\, Windsurf offers more than just code completion and ch at assistance – it's an agentic IDE built to enhance engineering flow. Th is talk explores how experienced engineers can leverage Windsurf's deep c ontextual awareness\, structured guidance\, and automated workflows to ta ckle sophisticated and complex tasks. We'll demonstrate practical strateg ies for accelerating feature development\, automating code maintenance an d reviews\, and ultimately freeing up cognitive load to focus on high-imp act engineering challenges. Learn how to move beyond basic AI assistance and truly partner with Windsurf to excel in your role. DTEND:20250604T111000 DTSTAMP:20260403T174839Z DTSTART:20250604T105500 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Mastering Engineering Flow with Windsurf UID:SZSESSION933684 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Tobin South\n\nEveryone is building MCP servers: from Slack integrations to personal data tools. They're good demos\, but not ready to turn into production. So\, what does it take to make MCP *enterp rise-ready?*\n\nWe're going to cover the end-to-end process of getting a hacky MCP server authenticated\, permissioned\, and secure. We'll talk ab out registries\, SSO\, audit logs\, agent identifiers\, autonomy for agen ts\, and oversight. Oh and we'll use MCP to buy some stuff.\n\nCome learn the stack needed to scale your MCP to the enterprise and some fun hacks along the way. DTEND:20250604T111000 DTSTAMP:20260403T174839Z DTSTART:20250604T105500 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:What does Enterprise Ready MCP mean? UID:SZSESSION933709 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Chris Kelly\n\nWhat's the role of vibe coding in a pr oduction-grade applications? Join Augment Code's Chris Kelly as he talks about the role of context in software engineering\, not code. DTEND:20250604T111000 DTSTAMP:20260403T174839Z DTSTART:20250604T105500 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Vibes won't cut it UID:SZSESSION936299 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Thibaut Gourdel\n\nWhile traditional RAG is effective \, it can struggle with complex relationships and reasoning across large knowledge bases. GraphRAG\, an advanced variant\, addresses these challen ges by leveraging knowledge graphs to enable deeper understanding and imp roved response accuracy. Learn how LLMs extract key entities and relation ships from your data to construct a graph structure\, and how the system uses graph traversal to find related entities and enrich prompts. Stay fo r a live demo showcasing these concepts in action. DTEND:20250604T111500 DTSTAMP:20260403T174839Z DTSTART:20250604T110000 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:GraphRAG: Integrating LLMs with Knowledge Graphs UID:SZSESSION933706 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Theodora Chu\n\nLearn more about the latest updates o n MCP and get ideas for what startups to build. DTEND:20250604T113500 DTSTAMP:20260403T174839Z DTSTART:20250604T111500 LOCATION:Yerba Buena Ballroom Salons 7-8: MCP SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:MCP Origins & RFS UID:SZSESSION947165 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Eric Simons\n\nTiny Teams are the future of how start ups are built\, and it all comes down to team culture\, decision making\, tooling choices\, and endless grit.\n\nIn this talk\, Eric will share th e high octane insights & learnings of how the 2nd fastest growing product in history _made it_ with a team of less than 15 people. DTEND:20250604T113500 DTSTAMP:20260403T174839Z DTSTART:20250604T111500 LOCATION:Yerba Buena Ballroom Salons 2-6: Tiny Teams SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Bolt.new: How we scaled $0-20m ARR in 60 days\, with 15 people UID:SZSESSION948608 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Eugene Yan\n\nRecommendation systems and search have long adopted advances in language modeling\, from early adoption of Word2 vec for embedding-based retrieval to the transformative impact of GRUs\, Transformers\, and BERT on predicting user interactions. Now\, the rise o f large language models (LLMs) is inspiring innovations in model architec ture\, scalable system designs\, and richer customer experiences.\n\nIn t his keynote\, we'll dive into cutting-edge industry applications of LLMs in recommendation and search systems\, exploring real-world implementatio ns and measurable outcomes. Join us for an look at current trends and an exciting vision of how LLM-driven techniques will shape the future of con tent discovery and intelligent search. DTEND:20250604T113500 DTSTAMP:20260403T174839Z DTSTART:20250604T111500 LOCATION:Golden Gate Ballroom A: LLM RecSys SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Recsys Keynote: Improving Recommendation Systems & Search in the A ge of LLMs UID:SZSESSION929337 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Mitesh Patel\n\nInterpreting complex information from unstructured text data poses significant challenges to Large Language Mo dels (LLM)\, with difficulties often arising from specialized terminology and the multifaceted relationships between entities in document architec tures. Conventional Retrieval Augmented Generation (RAG) methods face lim itations in capturing these nuanced interactions\, leading to suboptimal performance. In our talk\, we introduce a novel approach integrating Know ledge Graph-based RAG (GraphRAG) with VectorRAG\, designed to refine ques tion-answering (Q&A) systems for more effective information extraction fr om complex texts. Our approach employs a dual retrieval strategy that har nesses both knowledge graphs and vector databases\, enabling the generati on of precise and contextually appropriate answers\, thereby setting a ne w standard for LLMs in processing sophisticated data. DTEND:20250604T113500 DTSTAMP:20260403T174839Z DTSTART:20250604T111500 LOCATION:Golden Gate Ballroom B: GraphRAG SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:HybridRAG: A Fusion of Graph and Vector Retrieval to Enhance Data Interpretation UID:SZSESSION915992 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Joel Hron\n\nThis keynote will explore what it takes to move from basic generative assistants to fully agentic AI—systems that don’t just suggest but plan\, act\, and adapt—all within the structured\ , high-trust environments where professionals actually work. DTEND:20250604T113500 DTSTAMP:20260403T174839Z DTSTART:20250604T111500 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:From Copilot to Colleague: Building Trustworthy Productivity Agent s for High-Stakes Work UID:SZSESSION903524 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Alessio Fanelli\, Clay Bavor\n\nAs the amount of con sumer facing AI products grows\, the most forward leaning enterprises hav e created a new role: the AI Architect. These leaders are responsible for helping define\, manage\, and evolve their company's AI agent experience s over time.\n\nIn this session\, Clay Bavor (Cofounder of Sierra) will j oin Alessio Fanelli (co-host of Latent Space) in a fireside chat to share what it means to be an AI Architect\, success stories from the market\, and the future of the role. DTEND:20250604T113500 DTSTAMP:20260403T174839Z DTSTART:20250604T111500 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Rise of the AI Architect UID:SZSESSION941249 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Tanmai Gopal\n\nWe will review the different kinds of automation use-cases\, and the approach we used\, that will drive over a $100M of expected annual impact by deploying AI for business critical in itiatives. \n\nWe will discuss what kinds of automation initiatives becom e possible because of Gen AI. These were not tenable before because of th e amount of customization required per customer or per scenario\, and the kind of data involved in these workflows. Previously\, these workflows w ere driven manually which were both error prone and required expensive tr aining. \n\nTo replace or augment these manual business critical processe s\, automation _has_ to cross a very high bar of reliability. \n\nWe will share how we addressed the inherent non-determinism of Gen AI to create a predictable system that doesn’t have any surprising failure modes. We’l l also discuss how we worked with our existing data that was spread acros s various systems without an expensive centralisation and clean up effort . DTEND:20250604T113500 DTSTAMP:20260403T174839Z DTSTART:20250604T111500 LOCATION:Foothill C: Agent Reliability SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:AI Automation that actually works: $100M\, messy data\, zero surpr ises UID:SZSESSION916117 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Raiza Martin\n\nWe're in an awkward adolescent phase of AI product (design). But what if this chaotic moment is actually our g reatest opportunity? Enter the rebuilding revolution.\n\nIn this talk\, w e'll explore how the current state of AI interfaces offers a once-in-a-ca reer chance to rethink fundamental UX patterns\, with practical guidance on avoiding common pitfalls that plague first-generation AI products. \n\ nLearn how to balance technical constraints with user needs\, identify wh ich conventional wisdom to keep versus discard\, and ship AI experiences that actually delight users rather than frustrate them. DTEND:20250604T113500 DTSTAMP:20260403T174839Z DTSTART:20250604T111500 LOCATION:Foothill G 1&2: Product Management SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:[PM Keynote] Everything is ugly so go build something that isn't UID:SZSESSION925337 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Charles Frye\n\nEvery programmer needs to know a few things about hardware\, like processors\, memory\, and disks. Due to AI s ystems' extreme demand for mathematical processing power\, AI engineers n eed to know a few things about GPUs -- the world's most popular high-thro ughput mathematical co-processor.\n\nIn this talk\, I will explain the fu ndamental engineering constraints and design decisions that shape GPUs an d trace those up to some counter-intuitive facts about the performance ch aracteristics of AI systems\, with actionable insights for their deployer s and consumers. DTEND:20250604T113500 DTSTAMP:20260403T174839Z DTSTART:20250604T111500 LOCATION:Foothill F: Infrastructure SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:What every AI engineer needs to know about GPUs UID:SZSESSION933719 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Kwindla Kramer\, Sean DuBois\n\nSean DuBois of OpenA I and Pion\, and Kwindla Hultman Kramer of Daily and Pipecat\, will talk about why you have to design realtime AI systems from the network layer u p.\n\nMost people who build realtime AI apps and frameworks get it wrong. They build from either the model out or the app layer down. But unless y ou start with the network layer and build up\, you'll never be able to de liver realtime audio and video streams reliably. And perhaps even worse\, you'll get core primitives wrong: interruption handling\, conversation s tate management\, asynchronous function calling.\n\nSean and Kwin agree o n most things: old-school realtime systems people against the rest of the world. But they disagree on some important things\, too\, and will argue about those things live on stage. Do you need to give developers "thick" client-side realtime SDKs? Can you build truly great vendor neutral APIs ? (You'll be surprised which of them argues which side\, on that topic.) DTEND:20250604T113500 DTSTAMP:20260403T174839Z DTSTART:20250604T111500 LOCATION:Foothill E: Voice SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:[Voice Keynote] Your realtime AI is ngmi UID:SZSESSION915028 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: John Welsh\n\nWe recently released remote MCP support for both claude.ai and the Anthropic API. This talk will cover architect ural decisions we made in our implementation\, remote MCP authentication\ , supporting engineers who are building out agentic AI tools\, implementi ng custom internal transports\, and whatever else we can fit into 18 minu tes of your time. DTEND:20250604T115500 DTSTAMP:20260403T174839Z DTSTART:20250604T113500 LOCATION:Yerba Buena Ballroom Salons 7-8: MCP SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:What we learned from shipping remote MCP support at Anthropic UID:SZSESSION942943 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Sid Bendre\n\nIn this session\, I will be presenting a case study of Oleve's journey\, revealing how we've scaled a profitable multi-product portfolio with a tiny team. I'll walk you through the emer gence of "tiny teams\," our two-track engineering methodology that has be come our blueprint\, as well as an inside look at our technical alpha – s pecifically how we've engineered deterministic AI agents to deliver magic al and reliable consumer experiences to millions. You'll learn how we've built internal tools to grow leanly and created operating playbooks to sc ale operations without traditional headcount requirements. I'll also shar e our approach to scrappy infrastructure innovation and how our investmen t in internal tooling has served as a critical force multiplier. Finally\ , I'll give an overview of parts of the profitable portfolio playbook tha t keeps us lean\, adaptable\, and profitable across multiple product line s.\n\nStructure of talk:\n- the tiny teams revolution\n- the two-track en gineering approach\n- technical alpha: deterministic ai agents at scale\n - scrappy infrastructure innovation\n- internal tooling as a multiplier\n - the profitable portfolio playbook DTEND:20250604T115500 DTSTAMP:20260403T174839Z DTSTART:20250604T113500 LOCATION:Yerba Buena Ballroom Salons 2-6: Tiny Teams SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The New Lean Startup UID:SZSESSION914550 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Han Wang\, Mukuntha Narayanan\n\nPinterest Search in tegrates Large Language Models (LLMs) to enhance relevance scoring by com bining search queries with rich multimodal content\, including visual cap tions\, link-based text\, and user curation signals. A semi-supervised le arning framework enables scaling to large and multilingual datasets\, goi ng beyond English and limited human labels. These LLM-driven models are d istilled into efficient architectures for real-time serving\, with experi mental validation and large-scale deployment demonstrating substantial im provements in search relevance for Pinterest users worldwide. DTEND:20250604T115500 DTSTAMP:20260403T174839Z DTSTART:20250604T113500 LOCATION:Golden Gate Ballroom A: LLM RecSys SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:What We Learned from Using LLMs in Pinterest Search UID:SZSESSION932498 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Chin Keong Lam\n\n"Wisdom Discovery at Scale: Code Le ss KAG with n8n MultiAI Agents" DTEND:20250604T115500 DTSTAMP:20260403T174839Z DTSTART:20250604T113500 LOCATION:Golden Gate Ballroom B: GraphRAG SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Wisdom Discovery at Scale: Code Less KAG with n8n MultiAI Agents UID:SZSESSION914548 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Infant Vasanth\, Vaibhav Page\n\nInvestment Operatio ns teams are the backbone of asset and investment management firms. Their day-to-day work not only enables portfolio managers to respond swiftly t o market events but also ensures that complex\, unstructured data flows s eamlessly across the organization.\nIn this talk\, we introduce a modular \, Kubernetes-native AI framework purpose-built to scale custom Knowledge Apps across the enterprise. Designed with speed\, flexibility\, and comp liance in mind\, the framework empowers teams to launch production-grade document extraction applications in minutes instead of months\, unlocking new levels of automation and efficiency for investment management workfl ows.\nWe’ll also share how this framework has helped BlackRock streamline document extraction processes\, generate investment signals\, reduce ope rational overhead\, and accelerate the delivery of high-impact business u se cases—all while maintaining the robustness and control required in a r egulated industry. DTEND:20250604T115500 DTSTAMP:20260403T174839Z DTSTART:20250604T113500 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Accelerating Investment Operations: How BlackRock Builds Custom Kn owledge Apps at Scale. UID:SZSESSION904722 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Michael Albada\n\nGenerative AI has dramatically shor tened the distance between ideas and implementation\, enabling faster pro totyping and deployment than ever before. But while language models can s treamline individual tasks\, true transformation comes from combining the se capabilities into intelligent\, autonomous systems—AI agents.\n\nThis talk explores how to build and deploy foundation model-enabled agent syst ems that go beyond simple prompt chaining or chatbots. Drawing from real- world implementations and the latest research\, it offers a clear and pra ctical path to designing both single-agent and multi-agent systems capabl e of handling complex workflows with minimal oversight.\n\nAttendees will gain a deeper understanding of the core design principles behind agentic systems\, the architectural trade-offs involved in orchestrating multipl e agents\, and the strategies required to develop tailored solutions that enhance efficiency and innovation. Whether just beginning or scaling up\ , participants will leave with actionable insights to navigate the rapidl y evolving world of AI autonomy. DTEND:20250604T115500 DTSTAMP:20260403T174839Z DTSTART:20250604T113500 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Applications with AI Agents UID:SZSESSION907834 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Dexter Horthy\n\nHi\, I'm Dex. I've been hacking on A I agents for a while.\n\nI've tried every agent framework out there\, fro m the plug-and-play crew/langchains to the "minimalist" smolagents of the world to the "production grade" langraph\, griptape\, etc.\n\nI've talke d to a lot of really strong founders who are all building really impressi ve things with AI. Most of them are rolling the stack themselves. I don't see a lot of frameworks in production customer-facing agents.\n\nI've be en surprised to find that most of the products out there billing themselv es as "AI Agents" are not all that agentic. A lot of them are mostly dete rministic code\, with LLM steps sprinkled in at just the right points to make the experience truly magical.\n\nAgents\, at least the good ones\, d on't follow the "here's your prompt\, here's a bag of tools\, loop until you hit the goal" pattern. Rather\, they are comprised of mostly just sof tware.\n\nSo\, I set out to answer:\n\nWhat are the principles we can use to build LLM-powered software that is actually good enough to put in the hands of production customers?\n DTEND:20250604T115500 DTSTAMP:20260403T174839Z DTSTART:20250604T113500 LOCATION:Foothill C: Agent Reliability SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:12 Factor Agents - Principles of Reliable LLM Applications UID:SZSESSION914080 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: James Lowe\n\nSo you've built another cool demo. Now what? You have hype\, but not impact. You have kudos but no users. Ultima tely you have a demo\, but not a product.\n\nThe unique uncertainty of AI technology demands a new approach – beyond traditional product managemen t. You need an AI Product Manager. This talk explains why this role is es sential for building real AI products\, using real case studies from the incubator for Artificial Intelligence in the UK Government.\n\nMore impor tantly\, it reveals why your technical depth makes you uniquely suited to step into this critical leadership gap. Discover why could be the ideal candidate to be the AI Product Manager your product needs\, and how to st ep into that role. DTEND:20250604T115500 DTSTAMP:20260403T174839Z DTSTART:20250604T113500 LOCATION:Foothill G 1&2: Product Management SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Why your product needs an AI product manager\, and why it should b e you UID:SZSESSION914842 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Dr. Jasper Zhang\, PhD\n\nAI infrastructure today is caught in an endless cycle: build more data centers\, deploy more GPUs\, repeat. \n\nBut this approach is fundamentally flawed—expensive\, ineffic ient\, and environmentally unsustainable. \n\nIn this talk\, we will unpa ck why continuously expanding data centers masks deeper infrastructure in efficiencies\, and why leveraging a GPU marketplace to dynamically alloca te existing resources is essential. \n\nWe will explore practical use-cas es where companies scale GPU capacity flexibly\, startups gain affordable compute\, and idle GPUs are monetized\, enabling a future of sustainable and democratized AI infrastructure. DTEND:20250604T115500 DTSTAMP:20260403T174839Z DTSTART:20250604T113500 LOCATION:Foothill F: Infrastructure SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Why We Don’t Need More Data Centers UID:SZSESSION905305 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Peter Bar\n\nWhat does it take to go from blank page to live enterprise voice agent in 100 days?\n\nThat’s the challenge we to ok on with Fin Voice at Intercom. Enterprise customer service demands hig h-quality\, reliable voice interactions - but delivering that fast means wrestling with tough problems like latency\, hallucinations\, voice quali ty\, and answer accuracy.\n\nWe rapidly evaluated and integrated a full v oice stack - including transcription\, language model\, text-to-speech\, retrieval-augmented generation\, and telephony - while designing tools th at fit seamlessly into existing human support workflows.\n\nIn this sessi on\, I’ll share key lessons from our accelerated development of Fin Voice . We'll explore the technical and operational hurdles we faced\, the trad e-offs we made\, and how we built deployment and handover tools that work for customer service teams. You'll leave with insights into building AI- driven voice products that are both powerful and practical. DTEND:20250604T115500 DTSTAMP:20260403T174839Z DTSTART:20250604T113500 LOCATION:Foothill E: Voice SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Shipping an Enterprise Voice AI Agent in 100 Days UID:SZSESSION931123 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Harald Kirschner\n\nThe true power of Model Context P rotocol emerges when clients and servers collaborate across the full spec trum of the specification. This talk presents practical examples of how V S Code's comprehensive implementation of MCP transforms the capabilities of AI assistants\, making them more contextual\, efficient\, and user-fri endly. We'll showcase advanced features like dynamic tool discovery and w orkspace-aware roots\, demonstrating how they create experiences impossib le with standard tools integrations while confronting the reality gap bet ween MCP's theoretical potential and practical implementation challenges. DTEND:20250604T121500 DTSTAMP:20260403T174839Z DTSTART:20250604T115500 LOCATION:Yerba Buena Ballroom Salons 7-8: MCP SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Full Spectrum MCP: Uncovering Hidden Servers and Clients Capabilit ies UID:SZSESSION914489 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Hassan El Mghari\n\nIn this talk\, Hassan will go ove r how he builds open source AI apps that get millions of users like roomG PT.io (2.9 million users)\, restorePhotos.io (1.1 million users)\, Blinks hot.io (1 million visitors)\, and LlamaCoder.io (1.4 million visitors). H e'll go over his journey in AI\, demo some of the apps that he's built\, and dig into his tech stack and code to explain how he builds these apps from scratch. He’ll also go over how to market them and go over his top t ips and tricks for building great full-stack AI applications quickly and efficiently.\n\nThis talk will start from first principles and give you a glimpse into Hassan’s workflow of idea -> working app -> many users. Att endees should come out of this session equipped with the resources to bui ld impressive AI applications and understand some of the behind the scene s of how they’re built and marketed. This will hopefully serve as an educ ational and inspirational talk that encourages builders to go build cool things. DTEND:20250604T121500 DTSTAMP:20260403T174839Z DTSTART:20250604T115500 LOCATION:Yerba Buena Ballroom Salons 2-6: Tiny Teams SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Using OSS models to build AI apps with millions of users UID:SZSESSION911846 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Hamed Firooz\, Maziar Sanjabi\n\nWe will give a talk about our journey of building a foundation model for solving ranking and recommendation tasks DTEND:20250604T121500 DTSTAMP:20260403T174839Z DTSTART:20250604T115500 LOCATION:Golden Gate Ballroom A: LLM RecSys SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:360Brew LLM-based Foundation Model for Personalized Ranking and Re commendation UID:SZSESSION936205 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Sam Julien\n\nEnterprise knowledge bases are filled w ith "dense mapping\," thousands of documents where similar terms appear r epeatedly\, causing traditional vector retrieval to return the wrong vers ion or irrelevant information. When our customers kept hitting this wall with their RAG systems\, we knew we needed a fundamentally different appr oach.\n\nIn this talk\, I'll share Writer's journey developing a graph-ba sed RAG architecture that achieved 86.31% accuracy on the RobustQA benchm ark while maintaining sub-second response times\, significantly outperfor ming vector approaches.\n\nI'll survey the key techniques behind this per formance leap and why graph-based approaches excel with complex enterpris e information structures like product documentation\, financial documents \, and technical specifications that challenge traditional RAG systems. Y ou'll learn about using specialized LLMs to build semantic relationships\ , how compression techniques efficiently handle concentrated enterprise d ata patterns\, and how infusing key data points in the memory layer of th e LLM lowers hallucination.\n\nThe presentation will provide practical in sights into identifying when graph-based approaches make sense for your o rganization's specific data challenges\, helping you make informed archit ectural decisions for your next enterprise RAG system. DTEND:20250604T121500 DTSTAMP:20260403T174839Z DTSTART:20250604T115500 LOCATION:Golden Gate Ballroom B: GraphRAG SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:When Vectors Break Down: Graph-Based RAG for Dense Enterprise Know ledge UID:SZSESSION912811 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Donald Hruska\n\nAI agents are on the cusp of revolut ionizing work as we know it. The number of use cases software can tackle is set to explode as AI handles tasks requiring real judgment. But to cro ss the gap between an interesting AI prototype and an essential business tool\, you need agents built by developers with real guardrails and secur ity.\n\nThis means blending AI assistance with traditional coding in a mu ltimodal approach that maximizes efficiency and control. The future isn't about dropping in an LLM — it requires integrating any model\, any data\ , any system to deliver results. \n\nCompanies utilizing this approach ca n finally turn their slice of the $500B+ of total AI investment into real business results. \n DTEND:20250604T121500 DTSTAMP:20260403T174839Z DTSTART:20250604T115500 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How agents will unlock the $500B promise of AI UID:SZSESSION915616 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Amir Haghighat\n\nThis year kicked off with the DeepS eek-R1 news cycle breaking out of our AI Engineering bubble into the main stream tech and business world. Leaders at the highest levels of the larg est enterprises started asking how open source models could enhance and a ccelerate their AI strategy.\n\nOpen source models promise increased owne rship of AI systems: control over performance and price\, improved uptime and reliability\, better compliance\, and flexible hosting options. How are these promises playing out after months of implementation? In this ta lk\, I’ll draw on hundreds of conversations with AI leaders at enterprise companies to discuss what has — and hasn’t — changed about enterprise AI strategy in a world where open-source models compete on the frontier of intelligence. DTEND:20250604T121500 DTSTAMP:20260403T174839Z DTSTART:20250604T115500 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The Rise of Open Models in the Enterprise UID:SZSESSION915990 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Preeti Somal\n\nAs AI agents move from prototypes to production\, developers are running into new challenges with orchestratio n\, failure handling\, and infrastructure. This session will unpack lesso ns from teams already building real-world systems and share how to design for reliability from the start. DTEND:20250604T121500 DTSTAMP:20260403T174839Z DTSTART:20250604T115500 LOCATION:Foothill C: Agent Reliability SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Scaling AI agents without breaking reliability UID:SZSESSION913755 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Kenneth Auchenberg\n\nLearnings from building product s at Stripe and applying them in an AI native word DTEND:20250604T121500 DTSTAMP:20260403T174839Z DTSTART:20250604T115500 LOCATION:Foothill G 1&2: Product Management SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Shipping something to someone always wins UID:SZSESSION945392 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Alex Cheema\n\nThe hardware lottery: when a research idea wins because it is better suited to current hardware and software\, and not because it is universally superior.\n\nMachine learning researche rs often treat hardware as a fixed constraint and stop exploring beyond i t. Yet historically\, breakthroughs have come from algorithms that best a lign with the dominant hardware-software stack - neural networks being a classic example.\n\nIn this talk\, EXO Labs co-founder Alex Cheema will s hare recent algorithmic improvements for running large scale AI workloads on Apple Silicon.\n\nAlex will demonstrate how the EXO Framework enables inference\, fine-tuning\, and training of large ML models on Apple Silic on\, from the scale of one MacBook locally to clusters of colocated M3 Ul tra Mac Studios. DTEND:20250604T121500 DTSTAMP:20260403T174839Z DTSTART:20250604T115500 LOCATION:Foothill F: Infrastructure SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Large Scale AI on Apple Silicon using EXO UID:SZSESSION937137 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Brooke Hopkins\n\nThe reliability challenges facing v oice & chat AI deployment today mirror those that the autonomous vehicle industry confronted years ago. This talk explores how evaluation methodol ogies developed for self-driving cars can be transferred to create autono mous\, self-improving evaluation systems for conversational AI. Drawing f rom my experience building evaluation infrastructure at Waymo and now dev eloping Coval\, an enterprise-grade reliability platform for conversation al agents\, I'll demonstrate how systematic testing infrastructure is not just a technical requirement but a competitive advantage in the rapidly evolving AI landscape. DTEND:20250604T121500 DTSTAMP:20260403T174839Z DTSTART:20250604T115500 LOCATION:Foothill E: Voice SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:What we can learn from self driving in autonomous voice agents UID:SZSESSION915031 END:VEVENT BEGIN:VEVENT DESCRIPTION:https://latent.space/paperclub DTEND:20250604T130000 DTSTAMP:20260403T174839Z DTSTART:20250604T120000 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Latent Space Paper Club: AIEWF Special Edition — VIbhu Sapra UID:SZSESSION55c3d242-8eb6-4980-947b-4db467f4d044 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: David Cramer\n\nYou’ve heard a lot about MCP\, probab ly been given an AI mandate or two\, and are trying to figure out what’s real and what’s make believe. \n\nThis session will give practical advice for how you should be thinking about MCP\, the implementation pit falls\ , and where the speaker thinks things are going. DTEND:20250604T123500 DTSTAMP:20260403T174839Z DTSTART:20250604T121500 LOCATION:Yerba Buena Ballroom Salons 7-8: MCP SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:MCP isn’t good\, yet UID:SZSESSION947995 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Max Brodeur-Urbas\n\nAn overview of how Gumloop is sc aling automation across companies like Instacart\, Webflow and Shopify wi th less than 10 people. DTEND:20250604T123500 DTSTAMP:20260403T174839Z DTSTART:20250604T121500 LOCATION:Yerba Buena Ballroom Salons 2-6: Tiny Teams SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Gumloop's Path to be a 10 person unicorn UID:SZSESSION940118 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Daniel Chalef\n\nRAG is great for static knowledge re trieval—but terrible at memory. Vectorstore-based systems sold as memory lack relational and temporal awareness\, leading agents astray with outda ted or ambiguous information.\n\nDiscover how temporally-aware knowledge graphs—built by the open-source Graphiti framework—solve these limitation s. You’ll learn practical strategies to maintain precise\, context-rich m emory\, enabling agents to reason accurately about historical context and knowledge provenance. DTEND:20250604T123500 DTSTAMP:20260403T174839Z DTSTART:20250604T121500 LOCATION:Golden Gate Ballroom B: GraphRAG SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Stop Using RAG as Memory UID:SZSESSION915023 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Harrison Chase\n\nIt's easy to build a prototype of a n agent\, but hard to put an agent in production - especially in an enter prise setting. In this section\, will talk about three ingredients for bu ilding reliable agents in the enterprise. DTEND:20250604T123500 DTSTAMP:20260403T174839Z DTSTART:20250604T121500 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:3 ingredients for building reliable enterprise agents UID:SZSESSION937225 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Eric Hou\n\nYou’d never let a swarm of fresh interns ship to prod on day one—same deal with AI agents. Mentoring the Machine d ives into how acting like a tech lead (not just a user) turns those bots into real leverage. In this talk\, Eric will deliver practical advice for working with AI agents in the SDLC. He'll also preview how effective use of AI agents changes the calculus of software engineering at both a micr o and macro level.\n DTEND:20250604T123500 DTSTAMP:20260403T174839Z DTSTART:20250604T121500 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Mentoring the Machine UID:SZSESSION933545 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Anish Agarwal\, Matthew Schoenbauer\, Raaz Dwivedi\, Raj Agrawal\n\nSoftware is eating the world. AI is eating software. AI-p owered SWE means a whole lot more software is going to be written that po wers mission critical systems in the coming years\, with hardly any of it written by humans. Hence\, when these software systems inevitably break\ , it’s going to be next to impossible to troubleshoot them. Towards addre ssing this issue\, we’ll do a product launch of Traversal’s AI\, a signif icant step towards self-healing software systems. We will showcase how it is already used to autonomously troubleshoot production incidents in som e of the most complex enterprise environments. DTEND:20250604T123500 DTSTAMP:20260403T174839Z DTSTART:20250604T121500 LOCATION:Foothill C: Agent Reliability SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Production software keeps breaking and it will only get worse. He re’s how Traversal is fixing it. UID:SZSESSION915312 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Brian Balfour\n\nIf you’ve ever been blocked by vague specs\, shifting goals\, or chasing “vibes\,” things have only gotten me ssier in the age of AI. Everyone is obsessing over engineers doing PM wor k and PMs cranking out prototypes—but that skips the hardest question: Wh at should we build\, and why will it win? Today’s competitive landscape i s a knife-fight. When it’s trivial to ship “something\,” true differenti ation becomes brutally difficult.\n\nAt Reforge\, we built AI agents t hat analyze user feedback at scale\, perform real-time market analysis\, model feature impact\, and run continuous user research -- pushing us to rethink what "product work” actually looks like.\n\nIn this talk\, we’ll explore:\n\n- How to find a seam within the red ocean of incumbents\, wel l-funded upstarts\, and the horde of startups. \n- How to use real-time f eedback analysis\, competitive monitoring\, synthetic users\, AI-native r esearch to understand impact before it ships. \n- How to architect workfl ows where human intuition and machine intelligence ship product side by s ide. DTEND:20250604T123500 DTSTAMP:20260403T174839Z DTSTART:20250604T121500 LOCATION:Foothill G 1&2: Product Management SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Survive the AI Knife-Fight: Building Products That Win UID:SZSESSION914975 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Jack Dwyer\, Neil Dwyer\n\nThis is a talk that goes over our experience deploying Orpheus (Emotive\, Realtime TTS) to product ion. It will cover topics:\n\n- Latency and optimizations\n- High fidelit y voice clones w/ examples\n- Load balancing w/ multiple GPUs and multipl e LoRas DTEND:20250604T123500 DTSTAMP:20260403T174839Z DTSTART:20250604T121500 LOCATION:Foothill E: Voice SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Serving Voice AI at $1/hr: Open-source\, LoRAs\, Latency\, Load Ba lancing UID:SZSESSION937506 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250604T133000 DTSTAMP:20260403T174839Z DTSTART:20250604T123000 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Lunch UID:SZSESSION92cbfc54-95af-498b-9bb9-c896a8497624 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Keiji Kanazawa\, Nagkumar Arkalgud\n\nIn the age of autonomous AI agents\, ensuring their safety and reliability is paramount . But how can we proactively uncover vulnerabilities before they impact r eal-world scenarios? Enter Azure AI Evaluation SDK’s Red Teaming Agent—a cutting-edge tool designed to rigorously challenge your AI agents\, expos ing hidden risks and unexpected behaviors. This session will guide you th rough the powerful capabilities of Azure’s Red Teaming Agent\, demonstrat ing how it simulates adversarial scenarios\, stress-tests agentic decisio n-making\, and ensures your applications remain robust\, ethical\, and sa fe. You’ll learn practical techniques for systematically identifying weak nesses\, interpreting evaluation results\, and integrating safety checks into your development lifecycle. Join us to explore how embracing adversa rial testing not only mitigates risks but strengthens trust in your AI so lutions—keeping you ahead in the rapidly evolving landscape of responsibl e AI. DTEND:20250604T130500 DTSTAMP:20260403T174839Z DTSTART:20250604T124500 LOCATION:Nobhill C&D: Microsoft SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:AI Red Teaming Agent: Accelerate your AI safety and security journ ey with Azure AI Foundry UID:SZSESSION944039 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Sam Alba\n\nAI-powered agents promise faster\, easier software delivery\, but their unpredictable behavior often makes enginee rs hesitant to fully trust them with critical workflows. Sam Alba\, Co-fo under of Dagger (and previously co-creator of Docker)\, explains how team s can reliably integrate agents into their delivery pipelines by shifting how they structure and manage automation.\n\nHe'll share four practical strategies learned from real-world experience:\n\n1. Treat agents as work flow participants\, not isolated tools.\nStop using agents as disconnecte d scripts or IDE plugins. Treating them as first-class parts of your deli very process simplifies your architecture\, reduces hidden complexity\, a nd makes agent outcomes more predictable.\n\n2. Use many small agents ins tead of one big one.\nJust as software evolved from monoliths to microser vices\, software delivery benefits from smaller\, specialized agents with clearly defined responsibilities. Smaller agents are easier to understan d\, maintain\, and integrate.\n\n3. Define clear environments—the real le ver for reliability.\nInstead of chasing perfect prompts or models\, focu s on clearly defining the tools\, resources\, and permissions around your agents. Precisely controlling their environments makes agents behave con sistently and reliably.\n\n4. Design workflows for easy debugging and obs ervability.\nAgents will sometimes fail unexpectedly. Sam will share simp le\, effective ways to build clear tracing and observability into your wo rkflows from the start\, making debugging quicker and less frustrating.\n \nYou'll leave with practical\, immediately usable techniques that give y ou the confidence to trust AI agents in your software delivery pipelines. DTEND:20250604T130000 DTSTAMP:20260403T174839Z DTSTART:20250604T124500 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How to trust an agent with software delivery UID:SZSESSION933629 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Charles Frye\n\nOpen weights models and open source i nference servers have made massive strides in the year since we last got together at AIE World's Fair.\n\nWhere once we had only pirated LLaMA 2 w eights and Transformers\, we now have an embarrassment of riches. In fact \, we have too many choices! What's an AI engineer looking to self-host i nference to do?\n\nIn this session\, we'll share our benchmarking results from hundreds of runs across models\, frameworks\, and hardware. We'll a lso share tips and tricks from working with teams deploying LLM inference at scale. DTEND:20250604T130000 DTSTAMP:20260403T174839Z DTSTART:20250604T124500 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How fast are LLM inference engines anyway? UID:SZSESSION933716 END:VEVENT BEGIN:VEVENT DESCRIPTION:[Attendee-Only and Attendee-Led 10min lightning talks: see htt ps://crowdcomms.com/aiengineer25/qanda/41445 to vote/submit]\n\nWe compar ed popular agent frameworks against building from scratch and share the p ros and cons + a tier list of frameworks based on our testing! Dropping a blog post too! DTEND:20250604T131000 DTSTAMP:20260403T174839Z DTSTART:20250604T130000 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Agent frameworks are NOT all you need — Sanford Moskowitz UID:SZSESSION821181c4-e4d9-4950-8b2f-4a41d7e961af END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ado Kukic\n\nWe believe programming with AI is going through massive changes — again.\n\nTurns out the models yearn for the to ols and tokens. We hold them back if we make them ask before they can cha nge a file.\n\nGive them tools & tokens and everything changes: what we u se them for\, how we use them\, how many we run at the same time\, how th ey talk to each other\, how they talk to you\, what they even are...\n\nI t's all going to change.\n\nAnd with Amp\, we're embracing it.\n\nIf you want to find out where this is all going — come with us. DTEND:20250604T131500 DTSTAMP:20260403T174839Z DTSTART:20250604T130000 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Everything is changing UID:SZSESSION933636 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Mark Bain\n\nAre you ready to give your AI agents a m emory upgrade?\nJoin us for a fast-paced workshop exploring how memory ca n transform your agents.\n\nWhat You'll Do:\nLearn Leading Memory Solutio ns: Gain practical experience with open-source tools like Neo4j\, Cognee\ , Graphiti\, and Mem0.\nExplore Memory Types: Understand the theory behin d long-term\, short-term\, episodic\, semantic\, and other memory types.\ nDiscover Memory Benefits: Learn how memory improves recall\, contextual awareness\, and reasoning in autonomous agents.\nCompare Implementations: Get a snapshot of how different solutions implement memory—what’s built- in\, flexible\, and experimental. We'll also demonstrate GraphRAG memory solutions and a GraphRAG chat implemented with Google ADK.\nWhether you’r e working on AI copilots\, agentic workflows\, or research prototypes\, t his workshop will help you embed real memory into your AI stack. DTEND:20250604T134500 DTSTAMP:20260403T174839Z DTSTART:20250604T130000 LOCATION:Golden Gate Ballroom B: GraphRAG SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Make Your AI Agents Remember What They Do! UID:SZSESSION916143 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Anush Dsouza\, Julián Duque\n\nIn this workshop\, yo u’ll learn how to use Heroku Managed Inference and Agents to build agenti c applications. We’ll cover how to provision and deploy LLM models to you r app\, run untrusted code securely in Python\, Node.js\, Go\, and Ruby u sing built-in tools\, and use the Model Context Protocol (MCP) to connect tools and actions that extend your agents' capabilities. DTEND:20250604T134500 DTSTAMP:20260403T174839Z DTSTART:20250604T130000 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Agentic Applications with Heroku Managed Inference and Ag ents UID:SZSESSION939093 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Mason Egger\n\nAI Agents are distributed systems. Age nts need to connect and communicate with tools\, data repositories\, othe r agents\, etc.\, all over a network. Event-Driven Architecture is a comm on pattern for facilitating this connectivity\, using Events as the commu nication abstraction. However\, this pattern introduces complexities as w ell\, such as fragmented logic\, increased latency\, decreased observabil ity\, and more. But what if there were a way to get the benefits of Event -Driven Architecture without the complexities? Enter Durable Execution. I n this talk\, we'll discuss the pitfalls of Event-Driven Architecture\, h ow Durable Execution solves these issues\, and why Durable Execution\, no t Events\, is the correct abstraction for building AI Agents. DTEND:20250604T131500 DTSTAMP:20260403T174839Z DTSTART:20250604T130000 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Events are the Wrong Abstraction for Your AI Agents UID:SZSESSION933652 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Laurie Voss\n\nAt LlamaIndex we see a lot of agents b uilt every day\, and we've got a sense of what works and what doesn't. We 've distilled those learnings down into a series of patterns and best pra ctices for building real-world\, production agents\, and we're here to sh are them. You'll learn patterns for applying structure and guidance to fa mously nondeterministic LLMs and get concrete instruction on how to imple ment them. DTEND:20250604T131500 DTSTAMP:20260403T174839Z DTSTART:20250604T130000 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Effective agent design patterns in production UID:SZSESSION933621 END:VEVENT BEGIN:VEVENT DESCRIPTION:[Attendee-Only and Attendee-Led 10min lightning talks: see htt ps://crowdcomms.com/aiengineer25/qanda/41445 to vote/submit]\n\nTesting i s the bread and butter of creating sustainable procedural systems. With M L components though it is unclear how to build real confidence in how eac h component will be have. Evals help but there is a lot variability. Howe ver there have been significant advances in formal verification applied t o ML that make it possible to deeply validate performance of vision\, tab ular data and some language based systems. This quick talk covers how to start applying these methods. DTEND:20250604T132000 DTSTAMP:20260403T174839Z DTSTART:20250604T131000 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Unit testing and TDD for AI/ML: How to Apply formal methods to get us there — Steven Willmott UID:SZSESSIONed263d9b-5f8a-45d7-81c7-728a2115d1e2 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Cedric Vidal\n\nAs AI agents transition from experime ntal assistants to critical components of enterprise workflows\, reliably evaluating their performance becomes essential. But how do you systemati cally measure an AI agent’s capabilities\, contextual understanding\, and accuracy across diverse scenarios?\n\nIn this talk\, we'll dive deep int o the Azure AI Evaluation SDK\, an innovative tool designed to rigorously assess agentic applications. Learn how to create powerful evaluations us ing structured test plans\, scenarios\, and advanced analytics that pinpo int strengths and expose hidden weaknesses. Through practical examples an d real-world case studies\, you'll discover how companies are already lev eraging this SDK to enhance agent trustworthiness\, reliability\, and per formance.\n\nWhether you're developing conversational agents\, data-drive n decision-makers\, or autonomous workflow orchestrators\, this session e quips you with the techniques and insights needed to ensure your AI solut ions deliver exceptional value and exceed user expectations.""\n DTEND:20250604T133000 DTSTAMP:20260403T174839Z DTSTART:20250604T131000 LOCATION:Nobhill C&D: Microsoft SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Agentic Excellence: Mastering Evaluation of AI Agents with Azure A I Evaluation SDK UID:SZSESSION936818 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jonathon Belotti\n\nA peek under the hood of how we b uilt container checkpoints and restores to enable massively parallel agen tic workflows.\n\nNo one wants to wait on infrastructure. In this short t alk we’ll go through a demo and system design of container checkpoint/res tore\, which supports both burst autoscaling and agent branching for Moda l's serverless Functions and Sandboxes.\n\nCan you save a live container to a file? Can you save a live GPU? Come by the Modal booth to find out! DTEND:20250604T133000 DTSTAMP:20260403T174839Z DTSTART:20250604T131500 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Run 1000 branches of code with sandbox snapshotting UID:SZSESSION944044 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jon Peck\n\nAs developers\, we don't spend most of ou r time vibe-coding prototypes. More often\, we're adding features\, squas hing bugs\, and building tests for existing apps across a wide variety of services and technologies. Come learn how MCPs help GitHub Copilot to un tangle real engineering problems. By allowing agent mode to securely work with data sources\, testing tools\, infrastructure providers\, and even core DevOps tooling -- we can go beyond the hype\, and solve the actual e ngineering problems we face every day. DTEND:20250604T133000 DTSTAMP:20260403T174839Z DTSTART:20250604T131500 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Real-world MCPs in GitHub Copilot Agent Mode UID:SZSESSION936903 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Kshitij Grover\n\nYou’ve trained the model—now it’s t ime to train the business. This talk dives into the engineering behind pr icing systems that can evolve as fast as your AI stack.\n\nOrb CTO Kshiti j Grover will walk through how leading AI companies design infrastructure to support experimentation\, scale\, and real-world monetization constra ints.\n\nTopics include:\n- How to meter usage and map it to pricing with accuracy and auditability\n- Factoring in margins and underlying costs w hen designing pricing strategy\n- Handling complexity across motions: sel f-serve vs. enterprise\, pay-as-you-go vs. committed contracts\n- How to test pricing changes safely (and roll them back when needed)\n\nWhether y ou’re bootstrapping a pricing system from scratch or replacing a brittle V1\, you’ll leave with architectural patterns and mental models to make p ricing a first-class engineering concern.\n DTEND:20250604T133000 DTSTAMP:20260403T174839Z DTSTART:20250604T131500 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Revenue Engineering: How to Price (and Reprice) Your AI Product UID:SZSESSION933641 END:VEVENT BEGIN:VEVENT DESCRIPTION:[Attendee-Only and Attendee-Led 10min lightning talks: see htt ps://crowdcomms.com/aiengineer25/qanda/41445 to vote/submit]\n\nQuality a nd reliability determine the success or failure of real-world Chatbot App s. This talk will share a battle-tested framework for comprehensive end-t o-end chatbot testing. Topics covered: - Setting robust baseline criteria and policies for test evaluations - Implementing a steerable framework f or evaluations - Auto-generating multiple variants for exhaustive test co verage - Required analytics for test run reporting - Leveraging AI to mak e failures actionable DTEND:20250604T133000 DTSTAMP:20260403T174839Z DTSTART:20250604T132000 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Beyond Basic Evals: Production-Ready Testing for AI Chatbot Applic ations — Noah Moscovici UID:SZSESSION5444ad5f-4cb0-423b-8eff-238ffc56dcd5 END:VEVENT BEGIN:VEVENT DESCRIPTION:[Attendee-Only and Attendee-Led 10min lightning talks: see htt ps://crowdcomms.com/aiengineer25/qanda/41445 to vote/submit]\n\nif you we re on crypto Twitter\, u probably saw the slopbots pushing memecoins ever ywhere\, or heard some fund say 'our next billion users r gonna be ai age nts!'\, or smth to that effect. these are things you Can do\, i have a li ttle list of things you Should do. none of them involve memecoins or AI u sers. well almost none of them. i'd posit that blockchains and high-agenc y software both have very large problems that they can solve for each oth er\, have sum ideas\, will yap abt em. DTEND:20250604T134000 DTSTAMP:20260403T174839Z DTSTART:20250604T133000 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:What you should ACTUALLY use AI Agents for in web3 — yikesawjeez UID:SZSESSION5d4884ba-074a-4609-8f3e-bac551b680b2 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Stephen Chin\n\nAI models are getting tasked to do in creasingly complex and industry specific tasks where different retrieval approaches provide distinct advantages in accuracy\, explainability\, and cost to execute. GraphRAG retrieval models have become a powerful tool t o solve domain specific problems where answers require logical reasoning and correlation that can be aided by graph relationships and proximity al gorithms. We will demonstrate how an agent architecture combining RAG and GraphRAG retrieval patterns can bridge the gap in data analysis\, strate gic planning\, and retrieval to solve complex domain specific problems. DTEND:20250604T134500 DTSTAMP:20260403T174839Z DTSTART:20250604T133000 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Agentic GraphRAG: AI’s Logical Edge UID:SZSESSION933549 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Anushrut Gupta\n\nThe rapid progress in LLM capabilit y has not translated to increased reliability for business critical AI us e cases. The root-cause? Data is "not ready".\nConversational analytics d oesn't go beyond the analyst team because it's hard to verify if the gene rated queries are actually doing what they are supposed to.\nRAG based sy stems often fail to handle the breadth and depth of real world use-cases because it requires a prohibitive amount of preparation & maintenance of an underlying knowledge graph.\n\nAgentic AI systems need to hard-code sp ecific workflows to work reliably and end up looking more like software e ngineering with LLM calls instead of delivering on the promise of truly a gentic workflows.\n\nIn all of these failure modes\, the common culprit i s that the planning or reasoning done by the LLM fails to accurately capt ure the user's intent or the domain's context aka the lack of a well prep ared semantic data layer.\n\nEnterprise data is silo-ed and vastly varyin g levels of quality and the perfect "semantic layer" and "metadata" is a moving target. New data is continuously being created and business defini tions are rapidly changing and often entirely on-demand.\nIn this talk we 'll share how you can build and maintain a semantic data layer that is ma intained entirely by AI\, and show (with live examples) how that dramatic ally improves reliability of the AI system that needs dynamic access to d ata.\nWe'll demonstrate how this sufficiently augments existing RAG\, tex t-to-SQL and tool calling techniques and starts opening the door to relia ble AI deployments.\n DTEND:20250604T134500 DTSTAMP:20260403T174839Z DTSTART:20250604T133000 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:"Data readiness" is a myth: Make AI Reliabile with an Agentic Sema ntic Layer UID:SZSESSION933669 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Marco Casalaina DTEND:20250604T135500 DTSTAMP:20260403T174839Z DTSTART:20250604T133500 LOCATION:Nobhill C&D: Microsoft SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Agentic RAG: build a reasoning retrieval engine with Azure AI Sear ch UID:SZSESSION936908 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Tejashwa Tiwari\, Tejashwa Tiwari\n\nCome learn abou t why Windsurf is the premiere choice for engineers and enterprises alike in applications of AI for development. DTEND:20250604T140000 DTSTAMP:20260403T174839Z DTSTART:20250604T134500 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Windsurf & Wonders UID:SZSESSION933685 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Samuel Colvin\n\nEveryone is talking about agents\, a nd right after that\, they’re talking about agent-to-agent communications . Not surprisingly\, various nascent\, competing protocols are popping up to handle it.\n\nBut maybe all we need is MCP — the OG of GenAI communic ation protocols (it's from way back in 2024!).\n\nLast year\, Jason Liu g ave the second most watched AIE talk — “Pydantic is all you need”.\n\nThi s year\, I (the creator of Pydantic) am continuing the tradition by argui ng that MCP might be all we need for agent-to-agent communications.\n\nWh at I’ll cover:\n\n- Misusing Common Patterns: MCP was designed for deskto p/IDE applications like Claude Code and Cursor. How can we adapt MCP for autonomous agents?\n- Many Common Problems: MCP is great\, but what can g o wrong? How can you work around it? Can the protocol be extended to solv e these issues?\n- Monitoring Complex Phenomena: How does observability w ork (and not work) with MCP?\n- Multiple Competing Protocols: A quick run -through of other agent communication protocols like A2A and AGNTCY\, and probably a few more by June 😴\n- Massive Crustaceans Party: What might success look like if everything goes to plan? DTEND:20250604T142000 DTSTAMP:20260403T174839Z DTSTART:20250604T140000 LOCATION:Yerba Buena Ballroom Salons 7-8: MCP SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:MCP is all you need UID:SZSESSION911925 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Grant Lee\n\nSean reached out on X\, happy to do a ta lk on how to build a tiny team DTEND:20250604T142000 DTSTAMP:20260403T174839Z DTSTART:20250604T140000 LOCATION:Yerba Buena Ballroom Salons 2-6: Tiny Teams SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Tiny Teams UID:SZSESSION923914 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Yesu Feng\n\nDiscuss the foundation model strategy fo r personalization at Netflix based on this post https://netflixtechblog.c om/foundation-model-for-personalized-recommendation-1a0bd8e02d39 and rece nt developments. DTEND:20250604T142000 DTSTAMP:20260403T174839Z DTSTART:20250604T140000 LOCATION:Golden Gate Ballroom A: LLM RecSys SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:One model to rule recommendations: Netflix's Big Bet UID:SZSESSION932583 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Jesús Barrasa\, Michael Hunger\, Stephen Chin\n\nRAG has become one standard architecture component for GenAI applications to address hallucinations and integrate factual knowledge. While vector sea rch over text is common\, knowledge graphs represent a proven advancement by leveraging advanced RAG patterns to access and integrate interconnect ed factual information\, complementing the language skills of LLMs. This talk explores GraphRAG challenges\, implementation patterns\, and real-wo rld agentic examples with Google's ADK\, demonstrating how this approach delivers more trustworthy and explainable GenAI solutions with enhanced r easoning capabilities. DTEND:20250604T142000 DTSTAMP:20260403T174839Z DTSTART:20250604T140000 LOCATION:Golden Gate Ballroom B: GraphRAG SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Practical GraphRAG - Making LLMs smarter with Knowledge Graphs UID:SZSESSION915740 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ben Kus\n\nExplore the technical evolution of metadat a extraction at Box and how it shaped the foundation of our AI platform. We’ll walk through our transition to an agentic-first design—why it was n ecessary\, how we approached the rebuild\, challenges we encountered alon g the way\, and the advantages it unlocked. DTEND:20250604T142000 DTSTAMP:20260403T174839Z DTSTART:20250604T140000 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building an Agentic Platform UID:SZSESSION932429 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Nathan Wan\n\nWhile much of the AI innovation in heal thcare has centered on clinical and patient-facing applications\, Revenue Cycle Management (RCM) remains an underexplored yet critical domain. Giv en the growing financial pressures facing providers\, rethinking how heal thcare gets paid is essential to ensuring access and sustainability. The combination of which makes RCM an opportune area for AI disruption.\n\nTh is session explores how the combination of vast structured and unstructur ed data\, often rule-based workflows\, and direct financial opportunity t o drive meaningful outcomes. We’ll also share practical lessons from our journey evolving a traditional machine learning mindset to incorporate th e latest advances in Generative AI\, and how that shift is reshaping what 's possible in healthcare operations. DTEND:20250604T142000 DTSTAMP:20260403T174839Z DTSTART:20250604T140000 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:AI That Pays: Lessons from Revenue Cycle UID:SZSESSION916157 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Chau Tran\, Chau Tran\n\nWhile LLMs demonstrated imp ressive reasoning capabilities\, their out-of-the-box reasoning is akin t o hiring a brilliant but brand-new employee who doesn’t have the enterpri se context of “how things are done at this company”. In this talk\, I'll introduce “Workflow Search” as a paradigm to build enterprise-aware agent s that can balance predictability on common tasks\, and flexibility on un foreseen tasks.\n DTEND:20250604T142000 DTSTAMP:20260403T174839Z DTSTART:20250604T140000 LOCATION:Foothill C: Agent Reliability SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How to build Enterprise-aware agents UID:SZSESSION915338 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ben Stein\n\nA customer recently asked me: “Hey\, can I tag your AI agent in a Google Doc comment?”\n\nThe honest answer: I ha ve no idea! We never designed our agents to handle Google Doc comments\, but we tried it anyway… and it worked! The agent performed beautifully\, the customer was thrilled\, and I was left bewildered.\n\nWelcome to Prod uct Management for AI agents\, where roadmaps are fuzzy and we only learn the boundaries of our products after they’re released. When a product do esn’t follow predefined requirements but instead learns and improvises at runtime\, PMs must give up control and lean into uncertainty\, curiosity \, experimentation\, and fast feedback loops.\n\nThis talk is a field gui de for Product/Engineering teams navigating this new reality. We’ll cover how to write specs for affordances instead of features\, how to use AI e vals as a product development tool\, and how to perform User Acceptance T esting on undocumented emergent behavior. Most importantly\, we’ll explor e how to build trust with customers even when the answer is\, truthfully\ , “I don’t know.”\n\nIf you’re managing AI-native products in 2025 the sa me way you managed web apps in 2020\, you might find yourself A/B testing an agent that decided to go off and do C\, D\, and E all by themselves!\ n DTEND:20250604T142000 DTSTAMP:20260403T174839Z DTSTART:20250604T140000 LOCATION:Foothill G 1&2: Product Management SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Shipping Products When You Don’t Know What they Can Do UID:SZSESSION915648 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Dylan Patel\n\nAs AI reshapes the global balance of p ower\, the infrastructure behind it—chips\, data centers\, power\, and su pply chains—has become a new arena for geopolitical competition. This tal k explores how nations are racing to secure critical AI hardware\, contro l compute capacity\, and assert influence over the technologies and talen t that define the future. DTEND:20250604T142000 DTSTAMP:20260403T174839Z DTSTART:20250604T140000 LOCATION:Foothill F: Infrastructure SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:[Infra Keynote] Geopolitics of AI Infrastructure UID:SZSESSION928676 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jordan Dearsley\n\nWe’re entering a world where talki ng to machines feels as natural as talking to people. Voice is about to b ecome the dominant interface for technology - ambient\, always-on\, and h uman by default. To get there\, we need infrastructure that can orchestra te voice\, tools\, memory\, real-time reasoning and telephony. This talk explores the vision for voice and how we're making it work at scale. DTEND:20250604T142000 DTSTAMP:20260403T174839Z DTSTART:20250604T140000 LOCATION:Foothill E: Voice SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building the Voice-First Future: Omnipresent Agents that Listen\, Talk and Act UID:SZSESSION916079 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Alex Volkov\, Benjamin Eckel\, Steve Manuel\n\nAI En gineers deserve observable tools! \n\nMCP getting adoption means that les s and less of your agents code is running under your control\, and this h as DX and observability challenges\, let's fix that! \n\nJoin Alex Volkov from Weights & Biases and Steve Manual from mcp.run on this recap of the current state of MCP observability\, including the observable.tools init iative\, a recap of where the field stands and what to look forward to + a practical example of MCP tool usage evaluation framework from mcp.run! DTEND:20250604T144000 DTSTAMP:20260403T174839Z DTSTART:20250604T142000 LOCATION:Yerba Buena Ballroom Salons 7-8: MCP SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Observable tools - the state of MCP observability UID:SZSESSION915013 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Vikas Paruchuri\n\nWe scaled Datalab 5x this year - t o 7-figure ARR\, with customers that include tier 1 AI labs. We train cus tom models for document intelligence (OCR\, layout)\, with popular repos surya and marker.\n\nI'll talk about a new approach to building AI teams\ , including lessons I learned from Jeremy Howard\, and how we manage buil ding popular repos\, scaling revenue\, and training models with a tiny te am. DTEND:20250604T144000 DTSTAMP:20260403T174839Z DTSTART:20250604T142000 LOCATION:Yerba Buena Ballroom Salons 2-6: Tiny Teams SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Datalab: 40k stars\, 7-figure ARR\, SoTA models\, team of 3 UID:SZSESSION939097 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Tejaswi Tenneti\, Vinesh Gudla\n\n- Learn how Instac art uses cutting-edge LLMs to redefine search and product discovery. \n- Explore innovative solutions overcoming traditional search engine limitat ions for grocery shopping.\n- Discover how LLMs enhance user intent under standing and generate engaging content.\n- See practical applications of LLM technology to improve search relevance and user experience. DTEND:20250604T144000 DTSTAMP:20260403T174839Z DTSTART:20250604T142000 LOCATION:Golden Gate Ballroom A: LLM RecSys SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How Instacart transformed its search and discovery using an LLM-dr iven approach UID:SZSESSION929231 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ola Mabadeje\n\nTraditional ticketing and testing wor kflows for change management and network operations often operate indepen dently and lack critical real-world context and adaptive decision making capabilities. This fragmented approach results in delayed resolutions\, r epeated incidents\, escalations\, and dissatisfied stakeholders.\n\nThis session explores an innovative solution leveraging the synergy of natural language processing from IT Service Management (ITSM) systems\, Multi-ag ent reasoning\, and dynamic context derived from live knowledge network g raphs. Attendees will gain insights into an end-to-end architecture where natural language intents from ITSM tickets seamlessly integrate with exp erts AI agents for complex workflow tasks\, supported by continuous netwo rk knowledge graph ingestion pipelines.\n\nThrough a detailed production case study\, we will demonstrate how Agentic reasoning combined with dyna mic network knowledge graph contexts significantly improves critical vali dation and workflow interactions. The showcased results will highlight dr amatic improvements in ticket resolution efficiency\, accuracy of network testing\, and overall execution quality\, delivering tangible value to b oth technical teams and business stakeholders. DTEND:20250604T144000 DTSTAMP:20260403T174839Z DTSTART:20250604T142000 LOCATION:Golden Gate Ballroom B: GraphRAG SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Multi-Agent AI and Network Knowledge Graphs for Change Management and Network Testing UID:SZSESSION916063 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Kevin Madura\, Mo Bhasin\n\nIf software was eating t he world before\, knowledge work will soon be devoured by AI. In corporat e America there are thousands of hours spent on rote tasks every day by e mployees\, consultants\, and lawyers alike. But is AI really capable of r eplacing work in the real world yet? \nProductivity estimates from GenAI range from 1.5% (NBER) to 96% (☝ us! ️). In this talk we'll share war sto ries of where the answer is yes (and no) and how we reduced human time sp ent on tasks from days to minutes in high-impact situations. \nThe path f rom promise to actual product\, used in real world settings\, from our ex perience\, is still unmapped. Learn what we built\, how we built it - wit h code - and how we got stakeholder buy-in to deploy it. DTEND:20250604T144000 DTSTAMP:20260403T174839Z DTSTART:20250604T142000 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The Billable Hour is Dead\; Long Live the Billable Hour? UID:SZSESSION914890 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Denys Linkov\n\nYou've been given an AI mandate but d on't have additional headcount\, what next? Re-skilling\, up-skilling and team augmentation become essential to delivering on a new mandate. In th is talk we'll cover strategies to structure cross functional AI teams wit h domain experts\, software engineers and ML engineers. We'll cover key s kills and milestones that each traditional role can contribute to in uniq ue ways. DTEND:20250604T144000 DTSTAMP:20260403T174839Z DTSTART:20250604T142000 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Structuring a modern AI team UID:SZSESSION904822 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Sam Bhagwat\n\nOne current hot debate is should you m ake your top-level abstraction a ReAct type agent running in a loop? or s hould you make it a structured workflow graph?\n\nOpenAI is launching the ir new framework and throwing shade on workflow graph approaches\n\nTBH w e think this whole debate is kinda dumb. \n\nWe've seen a lot of folks be able to structure the problem in a way that a workflow graph makes a lot of sense. \n\nWe also see a ton of agents where you need to run the core bit in a loop for a long time.\n\nYou can also give your agents structur ed workflow graphs as a tool. You can use structured workflow graphs as a handoff mechanism between agents. What we've seen from the community is frankly that folks need to tinker with multiple approaches and combine pr imitives in interesting ways\n\nWe'll share a couple stories where teams ended up with workflow graph based approaches\, a couple where teams ende d up with agent based approaches\, and a couple where a blended approach made sense. DTEND:20250604T144000 DTSTAMP:20260403T174839Z DTSTART:20250604T142000 LOCATION:Foothill C: Agent Reliability SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Agents vs Workflows: Why Not Both? UID:SZSESSION914015 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Christopher Lovejoy\n\nVertical AI is a multi-trillio n-dollar opportunity. But you can't build a domain-expert application sim ply by grabbing the latest LLMs off-the-shelf: you need a system for codi fying latent insights from domain experts and using that to drive develop ment of your application.\n\nIn this talk\, we'll describe the system we' ve built at Anterior which has enabled us to achieve SOTA clinical reason ing and serve health insurance providers covering 50 million American liv es. We'll share:\n- how and why to encode domain-specific failure modes a s an ontology\n- a practical system for converting domain expertise into quantifiable eval metrics\n- how we structure work and collaboration betw een our clinicians\, engineer and PMs\n- our eval-driven AI iteration pro cess and how this can be adapted to any industry DTEND:20250604T144000 DTSTAMP:20260403T174839Z DTSTART:20250604T142000 LOCATION:Foothill G 1&2: Product Management SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Make your LLM app a Domain Expert: How to Build an LLM-Native Expe rt System UID:SZSESSION915738 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Kyle Kranen US\n\nYour model works! It aces the evals ! It even passes the vibe check! All that’s required is inference\, right ? Oops\, you’ve just stepped into a minefield:\n\n-Not low-latency enough ? Choppy experience. Users churn from your app. \n-Not cheap enough? You’ re losing money on every query.\n-Not high enough output quality? Your sy stem can’t be used for that application.\n\nA model and the inference sys tem around it form a “token factory” associated with a Pareto frontier— a curve representing the best possible trade-offs between cost\, throughpu t\, latency and quality\, outside of which your LLM system cannot be appl ied successfully. \n\nOutside of the Pareto frontier? You’re back to squa re one.\nThat is\, unless you’re able to change the shape of the Pareto f rontier.\n\nIn this session\, we’ll introduce NVIDIA Dynamo\, a datacente r-scale distributed inference framework as well as the bleeding-edge tech niques it enables to hack the Pareto frontier of your inference systems\, including:\n\n-Disaggregation - separating phases of LLM generation to m ake them more efficient\n-Speculation - predicting multiple tokens per cy cle\n-KV routing\, storage\, and manipulation - ensuring that we don’t re do work that has already been done\n-Pipelining improvements for agents - accelerating our workflows using information about the agent\n\nBy the e nd of the talk\, we’ll understand how the Pareto frontier limits where mo dels can be applied\, the intuition behind how inference techniques can b e used to modify it\, as well as the mechanics of how these techniques wo rk.\n DTEND:20250604T144000 DTSTAMP:20260403T174839Z DTSTART:20250604T142000 LOCATION:Foothill F: Infrastructure SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Hacking the Inference Pareto Frontier for Cheaper and Faster Token s Without Breaking SLAs UID:SZSESSION915471 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Tom Shapland\, PhD\n\nChatGPT Advanced Voice Mode isn ’t interrupting just you. Interruptions\, and turn-taking in general\, ar e unsolved problems for all Voice AI agents. Nobody likes being cut short – and people have much less patience for machines than they do for other humans. Turn-taking failures take many forms (e.g.\, the agent interrupt s the user\, the agent mistakes a cough for an interruption)\, and all of them lead to users immediately hanging up the phone.\n\nIn this talk\, w e use human conversation as a framework for understanding both today’s ap proaches to turn detection and where the field is headed. You’ll learn ab out how linguists think about turn detection in human dialogue\, what’s w orking (and what’s broken) in current methods\, and how we might build Vo ice AIs that interrupt you less than your human brother. DTEND:20250604T144000 DTSTAMP:20260403T174839Z DTSTART:20250604T142000 LOCATION:Foothill E: Voice SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Why ChatGPT Keeps Interrupting You UID:SZSESSION932493 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jan Curn\n\nThanks to MCP and all the MCP server dire ctories\, agents can now autonomously discover new tools and other agents . This lays down the foundation for the future agentic economy\, where bu sinesses will sell to autonomous agents (B2A) and eventually agents will sell to other agents (A2A).\n\nBut one key part is still missing: agents do not have a standard way to subscribe to external services and pay for them.\n\nIn this talk\, we’ll show how to give agents full autonomy to di scover and pay for new external MCP-enabled services\, even if those serv ices don’t support it\, using a little-known MCP server nesting capabilit y. We’ll also cover how to monetize AI agents and the B2A/A2A business mo dels.\n DTEND:20250604T150000 DTSTAMP:20260403T174839Z DTSTART:20250604T144000 LOCATION:Yerba Buena Ballroom Salons 7-8: MCP SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The rise of the agentic economy on the shoulders of MCP UID:SZSESSION926313 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Alex Duffy\n\nBenchmarks shape more than just AI mode ls—they shape our future. The things we choose to measure become self-ful filling prophecies\, guiding AI toward specific abilities and\, ultimatel y\, defining humanity’s evolving role in the AI era. Today’s benchmarks h ave propelled incredible progress\, but now we have an exciting opportuni ty: thoughtfully designing benchmarks around what genuinely matters to us —cooperation\, creativity\, education\, and meaningful human experiences. \n\nIn this talk\, we’ll explore how benchmarks function as powerful cult ural memes\, influencing not only technical outcomes but societal directi on. Drawing on practical examples we have seen at Every consulting in ind ustries like finance\, journalism\, education\, and even personally makin g AI play diplomacy. We’ll uncover what makes a benchmark impactful\, app roachable\, and inspiring. You’ll see our engaging new AI Diplomacy bench mark demo\, illustrating vividly how thoughtful evaluation design can exc ite both engineers and the wider community.\n\nYou’ll hopefully walk away inspired and equipped to define benchmarks intentionally\, helping steer AI toward outcomes that truly matter. DTEND:20250604T150000 DTSTAMP:20260403T174839Z DTSTART:20250604T144000 LOCATION:Yerba Buena Ballroom Salons 2-6: Tiny Teams SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Benchmarks Are Memes: How What We Measure Shapes AI—and Us UID:SZSESSION915974 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Devansh Tandon\n\nYouTube recommendations drive the m ajority of video watch time for billions of daily users. Traditionally po wered by large embedding models (LEMs)\, we're undertaking a fundamental shift: rebuilding our recommendation stack using foundation models like G emini. This talk dives into our engineering journey adapting general-purp ose LLMs (Gemini) for the highly specialized\, dynamic\, and massive-scal e task of YouTube recommendations.\n\nWe'll discuss: \n- SemanticID: crea ting a "language" for YouTube videos\, from our paper last year – Better Generalization with Semantic IDs: A Case Study in Ranking for Recommendat ions\n- Adapting Gemini checkpoints to understand SemanticID\n- Generativ e Video Retrieval with prompts \n\nThere’s a lot of attention on the LLM- led transformation of Search (with AI Overviews\, Perplexity\, ChatGPT-Se arch etc). However\, across large consumer apps\, it’s the recommendation systems & feeds that drive most consumer engagement\, not just search. T his talk is about the LLM-led transformation of recommendations & feeds – building a recommendation engine on top of Gemini. DTEND:20250604T150000 DTSTAMP:20260403T174839Z DTSTART:20250604T144000 LOCATION:Golden Gate Ballroom A: LLM RecSys SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Teaching Gemini to Speak YouTube: Adapting LLMs for Video Recommen dations to 2B+ DAU UID:SZSESSION906567 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Tom Smoker\n\nStructured Representations are pretty i mportant in the law\, where the relationships between clauses\, documents \, entities\, and multiple parties matter. Structured Representation mean s Structured Context Injection. Better Context\, Less Hallucinations. We walk through a couple of case studies of systems that we’ve built in prod uction for legal use-cases - from recursive contractual clause retrieval\ , to HITL legal reasoning news agents.\n\nYou'll gain insights into how s tructured representations significantly improve the effectiveness and rel iability of legal agents.\n DTEND:20250604T150000 DTSTAMP:20260403T174839Z DTSTART:20250604T144000 LOCATION:Golden Gate Ballroom B: GraphRAG SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Beyond Documents: Implementing Knowledge Graphs in Legal Agents UID:SZSESSION900332 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Eliza Cabrera\, Jeremy Silva\n\nAI across product\, GTM\, and strategy was a great approach in 2023\, but by now\, we all alr eady know that AI is disrupting the global landscape and how business get s done. Now is the time to stop chasing your competitors\, and letting th e technology lead your product strategy. There’s a better way to build th at will allow you to differentiate and keep pace.\n\nJoin AI product mana gers Eliza Cabrera and Jeremy Silva to learn how to crawl\, walk\, and ru n your way towards building dynamic products.\n DTEND:20250604T150000 DTSTAMP:20260403T174839Z DTSTART:20250604T144000 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Build Dynamic Products\, and Stop the AI Sideshow UID:SZSESSION915770 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Anoop Kotha\n\nHow to build production voice applicat ions and learnings from working with customers along the way DTEND:20250604T150000 DTSTAMP:20260403T174839Z DTSTART:20250604T144000 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Effective Voice Agents UID:SZSESSION915067 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Itamar Friedman\n\nEveryone wants to do Vibe Code\, e ven large Enterprises. But how can we ensure that the generated code is w ell-grounded with the dev team's code and software development standards? In this talk\, Itamar will present how to use various tools and agents\, including MCP and A2A\, to achieve precisely that. DTEND:20250604T150000 DTSTAMP:20260403T174839Z DTSTART:20250604T144000 LOCATION:Foothill C: Agent Reliability SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Vibe Coding\, with Confidence UID:SZSESSION916115 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Tom Moor\n\nLearn how we're evolving Linear into an o perating system for engineering teams to ship product with agents as a fi rst class citizen. DTEND:20250604T150000 DTSTAMP:20260403T174839Z DTSTART:20250604T144000 LOCATION:Foothill G 1&2: Product Management SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building the platform for agent coordination UID:SZSESSION940848 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Robert Wachen\n\nCurrent AI inference systems rely on brute-force scaling—adding more GPUs for each user—creating unsustainabl e compute demands and spiraling costs. Real-time use cases are bottleneck ed by their latency and costs per user. In this talk\, AI hardware expert and founder Robert Wachen will break down why the current approach to in ference is not scalable\, and how rethinking hardware is the only way to unlock real-time AI at scale. DTEND:20250604T150000 DTSTAMP:20260403T174839Z DTSTART:20250604T144000 LOCATION:Foothill F: Infrastructure SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Flipping the Inference Stack: Why GPUs Bottleneck Real-Time AI at Scale UID:SZSESSION912986 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Kwindla Kramer\, Shrestha Basu Mallick\n\nThe Gemini Live API GA is now powered by Google's best cost-effective thinking mod el Gemini 2.5 Flash. We will do a deep dive on the capabilities that the Gemini Live API combined with Pipecat unlock for devs with special focus on session management\, turn detection\, tool use (including async functi on calls)\, proactivity\, multilinguality and integration with telephony and other infra. We will demo some of the more innovative capabilities. W e will also talk through some customer use cases - especially how custome rs can use Pipecat to extend these realtime multimodal capabilities to cl ient side applications such as customer support agents\, gaming agents\, tutoring agents etc. In addition\, we also have an experimental version o f the Live API powered by with Google's native audio offering that can be tried in an experimental capacity . This experimental model can communi cate with seamless\, emotive\, steerable\, multilingual dialogue and enha nces use cases where more natural voices can be a big differentiator. DTEND:20250604T150000 DTSTAMP:20260403T174839Z DTSTART:20250604T144000 LOCATION:Foothill E: Voice SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Milliseconds to Magic: Real‑Time Workflows using the Gemini Live A PI and Pipecat UID:SZSESSION933596 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Dat Ngo\n\nModel-Context-Protocol (MCP) has become th e “USB-C for AI.” But when a request jumps across tools\, clouds and lang uages\, debugging becomes tricky. Where does your latency spike? What bro ke your prompt? In this lightning demo\, we show how open source Arize Ph oenix + OpenInference propagate OpenTelemetry context from an MCP client ➜ server ➜ Claude chain\, letting you watch every span in real time\, spo t bottlenecks\, and replay bad traces in seconds. You’ll leave knowing ho w to gain full-stack visibility for your MCP Apps without any custom code or vendor lock-in.\n DTEND:20250604T151000 DTSTAMP:20260403T174839Z DTSTART:20250604T150000 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How To Implement End-to-End Tracing for MCP Client-Server Applicat ions UID:SZSESSION946084 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250604T154500 DTSTAMP:20260403T174839Z DTSTART:20250604T150000 LOCATION:Golden Gate Foyer SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Afternoon Break UID:SZSESSIONa8395f63-f500-42ec-9fba-ef91cece2295 END:VEVENT BEGIN:VEVENT DESCRIPTION:[Attendee-Only and Attendee-Led 10min lightning talks: see htt ps://crowdcomms.com/aiengineer25/qanda/41445 to vote/submit]\n\nImagine i f product managers had access to the latest vibecoding tools to create UI s\, integrate APIs\, and test workflows on their own. What if they could transition from concept to wireframe and user testing in just a few days instead of several weeks? In this talk\, I will share how we successfully onboarded our product managers onto low-code vibecoding platforms. We re duced prototyping time and empowered non-technical team members to delive r DTEND:20250604T152500 DTSTAMP:20260403T174839Z DTSTART:20250604T151500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How we turned PMs into Vibecoders — Nicolas Grenie UID:SZSESSIONcb7b3069-14f6-4806-afcf-1e97f0175031 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jim Bennett\n\nLLM agents often drift into failure wh en prompts\, retrieval\, external data\, and policies interact in unpredi ctable ways. This session introduces a repeatable\, metric-driven framewo rk for detecting\, diagnosing\, and correcting these undesirable behavior s in agentic systems at production scale. DTEND:20250604T153000 DTSTAMP:20260403T174839Z DTSTART:20250604T151500 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Taming Rogue AI Agents with Observability-Driven Evaluation UID:SZSESSION936894 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Philipp Krenn\n\nEvery vector database out there is b oth faster and slower than any other competitor — if you believe all the benchmarketing out there.\nLet's turn the marketing into useful benchmark s that actually help you:\n1. How not to benchmark (spoiler: don’t trust the glossy charts).\n2. What’s uniquely tricky about benchmarking vector search.\n3. How to build meaningful benchmarks tailored to your use case. \n\nPS: Yes\, you will have to get your hands dirty. Never believe a benc hmark that you haven't tweaked yourself. DTEND:20250604T153000 DTSTAMP:20260403T174839Z DTSTART:20250604T151500 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Vector Search Benchmark[eting] UID:SZSESSION933625 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jonathan Larson DTEND:20250604T153000 DTSTAMP:20260403T174839Z DTSTART:20250604T151500 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:GraphRAG methods to create optimized LLM context windows for retri eval UID:SZSESSION936906 END:VEVENT BEGIN:VEVENT DESCRIPTION:[Attendee-Only and Attendee-Led 10min lightning talks: see htt ps://crowdcomms.com/aiengineer25/qanda/41445 to vote/submit]\n\nMCP trace s can get lost and disjointed in the data transport between the client an d the server. Even when connected these traces can remain raw and hard to read unless tracing is diligently built. Both these flaws make MCP obser vability a nonstarter. This talk will demonstrate how to build beautiful continuous traces in a single line of code! (the only constrain is that i t must use the MCP library under the hood) DTEND:20250604T153500 DTSTAMP:20260403T174839Z DTSTART:20250604T152500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:MCP End 2 End Traces For Fully Observable Agents — Sebastian Sosa UID:SZSESSION32855e25-6c36-4ddc-8ba2-213698d3e6d5 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Dat Ngo\n\nAs LLM-powered products become more sophis ticated\, the need for scalable\, reliable evaluation pipelines has never been more critical. This session dives deep into advanced LLM evaluation strategies that move beyond toy benchmarks and toward real-world product ion impact.\n\nWe’ll explore how to architect and implement evaluation pi pelines that work across both online and offline environments—reducing de v complexity and accelerating iteration. The session will cover:\n\n- LLM -as-a-judge frameworks\n- Human-in-the-loop evaluation\n- How hybrid appr oaches unlock more robust and nuanced performance assessments\n\nWe’ll br eak down technical architectures\, share real implementation patterns\, a nd examine trade-offs between evaluation techniques to help engineers mak e informed choices.\nWhether you’re building from scratch or refining exi sting workflows\, this talk offers practical strategies for crafting effi cient\, scalable\, and accurate eval pipelines tailored to custom LLM pro ducts. DTEND:20250604T154500 DTSTAMP:20260403T174839Z DTSTART:20250604T153000 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Engineering Better Evals: Scalable LLM Evaluation Pipelines That W ork UID:SZSESSION936204 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Matthias Loibl\n\nPolar Signals Continuous Profiling for GPUs extends our industry-leading continuous profiling platform to pr ovide deep\, always-on visibility into your GPU workloads.\n\nNow you can see exactly how your GPUs are being utilized millisecond by millisecond. Our solution helps you move from guesswork to data-driven optimization. DTEND:20250604T154500 DTSTAMP:20260403T174839Z DTSTART:20250604T153000 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Maximize GPU Efficiency with Continuous Profiling for GPUs UID:SZSESSION937936 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Mani Khanuja\n\nAs organizations seek to harness the ir proprietary data while maintaining security and compliance\, Amazon B edrock provides a comprehensive framework for building tailored AI appli cations. Using Amazon Bedrock Knowledge Bases and Amazon Bedrock Data Au tomation\, organizations can create AI solutions that truly understand t heir unique business context\, terminology\, and requirements. Combined with Amazon Bedrock Guardrails\, these capabilities enhance the accuracy and relevance of AI-generated responses\, while ensuring that sensitive information remains protected within the organization's control - enabl ing businesses to build secure and compliant enterprise-grade generative AI solutions that accelerate time to value. DTEND:20250604T154500 DTSTAMP:20260403T174839Z DTSTART:20250604T153000 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Data is Your Differentiator: Building Secure and Tailored AI Syste ms UID:SZSESSION933605 END:VEVENT BEGIN:VEVENT DESCRIPTION:[Attendee-Only and Attendee-Led 10min lightning talks: see htt ps://crowdcomms.com/aiengineer25/qanda/41445 to vote/submit] DTEND:20250604T154500 DTSTAMP:20260403T174839Z DTSTART:20250604T153500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Hallway Track 1: Walk-up Talk if time permits UID:SZSESSION85bc0a79-8240-4470-84ef-1884884a89f1 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250604T160000 DTSTAMP:20260403T174839Z DTSTART:20250604T154500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Keynote Doors UID:SZSESSION6388169d-662b-4bd9-bbd4-ad439d46377b END:VEVENT BEGIN:VEVENT DESCRIPTION:Highlights from a great GraphRAG track. DTEND:20250604T160500 DTSTAMP:20260403T174839Z DTSTART:20250604T160000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Closing thoughts on Agentic GraphRAG UID:SZSESSION16a9de0c-13f7-4c1e-a911-b476f0eec737 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Antje Barth\n\nLet's explore practical strategies fo r building and scaling agents in production. Discover how to move from l ocal MCP implementations to cloud-scale architectures and how engineerin g teams leverage these patterns to develop sophisticated agent systems. Expect a mix of demos\, use case discussions\, and a glimpse into the fu ture of agentic services! DTEND:20250604T162500 DTSTAMP:20260403T174839Z DTSTART:20250604T160500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Agents at Cloud-Scale UID:SZSESSION933612 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Kevin Hou\n\nabstract tbd DTEND:20250604T164500 DTSTAMP:20260403T174839Z DTSTART:20250604T162500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Windsurf everywhere\, doing everything\, all at once UID:SZSESSION933675 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Greg Brockman\, swyx .\n\nGreg Brockman's career and advice for AI Engineers DTEND:20250604T172500 DTSTAMP:20260403T174839Z DTSTART:20250604T164500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:#define AI Engineer UID:SZSESSION936006 END:VEVENT BEGIN:VEVENT DESCRIPTION:Our After Party will take place in the Expo Hall and Grand Ass embly DTEND:20250604T190000 DTSTAMP:20260403T174839Z DTSTART:20250604T171500 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The Tollbit afterparty UID:SZSESSION2ade624f-6ade-4575-b62e-6c9c1c37fbb8 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ben Dunphy DTEND:20250604T173500 DTSTAMP:20260403T174839Z DTSTART:20250604T172500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Welcome to the Tollbit Afterparty UID:SZSESSION946658 END:VEVENT BEGIN:VEVENT DESCRIPTION:Organize your own in the app/AIE slack\, or sign up at https:/ /www.ai.engineer/#events DTEND:20250604T224000 DTSTAMP:20260403T174839Z DTSTART:20250604T190000 LOCATION:Atrium: Event Hub SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Community Meetups (Jun 4) UID:SZSESSIONc67d8e1a-a720-4da5-a017-37f80442a684 END:VEVENT BEGIN:VEVENT DESCRIPTION:Registration is open from 7:00am-3:00pm in the Event Hub. DTEND:20250605T150000 DTSTAMP:20260403T174839Z DTSTART:20250605T070000 LOCATION:Atrium: Event Hub SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Registration UID:SZSESSION972e15da-8aa9-4b84-abf8-7c6c9d9502cb END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250605T095500 DTSTAMP:20260403T174839Z DTSTART:20250605T071500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Continental Breakfast UID:SZSESSION3a681f8f-d56e-4261-8cc9-048ae26c50c8 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250605T084500 DTSTAMP:20260403T174839Z DTSTART:20250605T074500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Rehearsals/Tech Check UID:SZSESSIONa32c03d2-4754-48a2-a0fa-4e520b1564bc END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250605T090000 DTSTAMP:20260403T174839Z DTSTART:20250605T084000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Keynote Doors UID:SZSESSION8ea2084c-0e8f-4a58-9a52-5f022c0705f3 END:VEVENT BEGIN:VEVENT DESCRIPTION:Laurie and Ben. DTEND:20250605T090500 DTSTAMP:20260403T174839Z DTSTART:20250605T090000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Kickoff + Special Announcement UID:SZSESSIONad58f33b-8b96-440e-8f0a-6f45fad1e1a6 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Logan Kilpatrick\n\nOver the last year\, Google and G emini models have shown rapid progress across all dimensions (model\, pro duct\, etc). Let's highlight all the work that has happened\, how we got the worlds best models\, and where we are going next (across both the mod el landscape and out AI products). DTEND:20250605T092500 DTSTAMP:20260403T174839Z DTSTART:20250605T090500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:A year of Gemini progress + what comes next UID:SZSESSION935461 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jack Rae\n\nProgress towards general intelligence has been marked by identifying fundamental intelligence bottlenecks within e xisting models and developing solutions that improve the architecture or training objective. From this perspective\, we discuss our work on Thinki ng in Gemini as a solution to a bottleneck in test-time compute. We will discuss recent progress in Thinking both from the benefit of capability a nd steerability\, and discuss where our models are headed. DTEND:20250605T094500 DTSTAMP:20260403T174839Z DTSTART:20250605T092500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Thinking Deeper in Gemini UID:SZSESSION947233 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Manu Goyal\n\nAn introduction to the evals track DTEND:20250605T095000 DTSTAMP:20260403T174839Z DTSTART:20250605T094500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Why should anyone care about Evals? UID:SZSESSION942167 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Solomon Hykes\n\nAI agents promise breakthroughs but often deliver operational chaos. Building reliable\, deployable systems w ith unpredictable LLMs feels like wrestling fog – testing outputs alone i s insufficient when the underlying workflow is opaque and flaky. How do w e move beyond fragile prototypes?\n\nThis talk\, from the creator of Dock er\, argues the solution lies *outside* the model: engineering **reproduc ible execution workflows** built on rigorous architectural discipline. Le arn how **containerization**\, applied not just to deployment but to *eac h individual step* of an agent's workflow\, provides the essential **isol ation and environmental consistency** needed.\n\nDiscover how combining t his granular container approach with patterns like immutable state manage ment allows us to **contain agent chaos**\, unlock effective testing\, si mplify debugging\, and bring essential control and predictability back to building powerful AI agents you can actually ship with confidence. DTEND:20250605T101000 DTSTAMP:20260403T174839Z DTSTART:20250605T095000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Containing Agent Chaos UID:SZSESSION916116 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jesse Han\n\nWe're at an inflection point where AI ag ents are transitioning from experimental tools to practical coworkers. Th is new world will demand new infrastructure for RL training\, test-time s caling\, and deployment. This is why Morph Labs developed Infinibranch la st year\, and we are excited to finally unveil what's next. DTEND:20250605T103000 DTSTAMP:20260403T174839Z DTSTART:20250605T101000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The infrastructure for the singularity UID:SZSESSION916189 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250605T111500 DTSTAMP:20260403T174839Z DTSTART:20250605T103000 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Morning Break UID:SZSESSIONcfb161c6-218d-44da-9ad2-36fccd701e40 END:VEVENT BEGIN:VEVENT DESCRIPTION:[last round of Attendee-Led 10min lightning talks]\n\nI will s howcase how I got tired of waiting for an AI assisted/no spoiler book rea ding experience and built my own. Check 30s video at youtu.be/JjwnYqy668M or go to demo book at https//bookgenius.net\n\nOpen Sourcing! DTEND:20250605T105500 DTSTAMP:20260403T174839Z DTSTART:20250605T104500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Books reimagined: use AI to create new experiences for things you know — Lukasz Gandecki UID:SZSESSION0f46436a-c182-4b85-9e03-f83040ae1773 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Suman Debnath\n\nBuilding AI agents used to require c omplex orchestration\, extensive scaffolding\, and months of tuning. With Strands Agents\, an open source SDK from AWS. You can now build\, test\, and deploy intelligent agents in just a few lines of code. This session introduces the model-driven approach behind Strands\, where a model\, a p rompt\, and a set of tools are all you need to create powerful\, producti on-ready agents. Learn how Strands leverages modern foundation models to handle reasoning\, tool use\, and reflection\, reducing development time from months to days. DTEND:20250605T110000 DTSTAMP:20260403T174839Z DTSTART:20250605T104500 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Introducing Strands Agents\, an Open Source AI Agents SDK UID:SZSESSION933603 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Kenneth DuMez\n\nLearn the secrets behind the workflo ws that engineers at the fastest moving companies in the world are using to build software for billions of users worldwide. This workshop will cov er a comprehensive overview of how to leverage generative AI to write cod e\, how to stack and submit these pull requests\, and finally how to use AI to review them. DTEND:20250605T110000 DTSTAMP:20260403T174839Z DTSTART:20250605T104500 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The fastest software dev workflow in the world: AI meets stacked d iffs UID:SZSESSION933462 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jesús Barrasa\n\nYou're trying to guide how your agen ts think and act. Code-orchestrated workflows are too rigid\, but LLMs ch arting their own course feel too chaotic. When you need a middle ground\, it’s time to reach for the secret weapon: ontologies. These graph-shaped fragments of actionable knowledge can fill in critical gaps.\n\nIn this talk\, we’ll explore together how ontologies bring structure\, semantics\ , and sanity to GenAI-powered applications. You’ll learn when they’re use ful\, how to apply them\, and what kinds of problems they help solve. Thr ough practical examples\, we’ll show how ontologies (1) guide knowledge g raph construction\, (2) add a semantic layer for more efficient and accur ate retrieval (GraphRAG)\, and (3) encode domain logic you don’t want to leave up to the LLM. DTEND:20250605T110000 DTSTAMP:20260403T174839Z DTSTART:20250605T104500 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Why Your Agent’s Brain Needs a Playbook: Practical Wins from Using Ontologies UID:SZSESSION933646 END:VEVENT BEGIN:VEVENT DESCRIPTION:[last round of Attendee-Led 10min lightning talks]\n\nI vibe c oded a college counselor app around Claude and Gemini\, here's what it to ok to pull off DTEND:20250605T110500 DTSTAMP:20260403T174839Z DTSTART:20250605T105500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Thick Wrappers in Practice — Dan Mason UID:SZSESSION8aeda13b-3b22-4f16-8435-65406eca79b4 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Mark Backman\n\nVoice AI agents today can conduct nat ural\, human-like conversations and perform a wide variety of tasks: cust omer support\, lead qualification\, healthcare patient intake\, market re search\, and more.\n\nToday's best voice agents combine: realtime respons iveness\, open-ended conversational intelligence\, reliable instruction f ollowing\, and flexible integration with existing back-end systems.\n\nLe arn how to build state of the art voice agents using Pipecat's open sourc e\, vendor neutral tooling. You can deploy Pipecat agents to your own inf rastructure or to Pipecat Cloud.\n\nPipecat is used and supported by team s at NVIDIA\, AWS\, Google DeepMind\, OpenAI\, and hundreds of other comp anies. DTEND:20250605T111500 DTSTAMP:20260403T174839Z DTSTART:20250605T110000 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Pipecat Cloud: Enterprise Voice Agents Built On Open Source UID:SZSESSION933589 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Chris Hernandez\n\nProduct leaders see AI possibiliti es. Operations teams see implementation chaos. That disconnect can kill p romising AI features before they ever reach users.\n\nIn this session\, C hris Hernandez and Jeremy Silva share an integrated framework that bridge s product strategy and operational reality. You'll learn how they transfo rmed fragmented AI workflows into a unified approach—from prototyping and prompt testing to human review loops and model benchmarking.\n\nWe’ll ex plore how to build evaluation systems that satisfy both technical and bus iness stakeholders\, create effective HITL processes from day one\, and u se QA as a strategic enabler of generative AI quality. Most importantly\, we’ll show how product and operations can move beyond friction—working t ogether to deliver AI features that scale responsibly and ship faster\, w ith confidence. DTEND:20250605T111500 DTSTAMP:20260403T174839Z DTSTART:20250605T110000 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The Build-Operate Divide: Bridging Product Vision and AI Operation al Reality UID:SZSESSION933622 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Frank Liu\n\nIn this talk\, we examine the state-of-t he-art in AI-powered search and retrieval. We detail techniques for enhan cing performance beyond base embedding models\, including hybrid search\, reranking strategies\, query decomposition and document enrichment\, the use of domain-specific and fine-tuned embeddings\, custom data processin g pipelines (ETL)\, and contextualized chunking methods. DTEND:20250605T111500 DTSTAMP:20260403T174839Z DTSTART:20250605T110000 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The State of AI-Powered Search and Retrieval UID:SZSESSION933689 END:VEVENT BEGIN:VEVENT DESCRIPTION:[last round of Attendee-Led 10min lightning talks]\n\nAI broke recruitment - how to think about hiring for AI-enabled engineers in the era of AI cheating agents and AI customised resumes\nBeth Glenfield DTEND:20250605T111500 DTSTAMP:20260403T174839Z DTSTART:20250605T110500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:AI broke recruitment — Beth Glenfield UID:SZSESSIONa977c94a-9d54-4b46-9615-2283a96b6e60 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Deepsha Menghani\, Rossella Blatt Vital\n\nWhat does it really take to move a modern SaaS company from AI experimentation to becoming truly AI-first?\n\nAt Sprout Social\, we’re in the midst of that transformation—rearchitecting strategy\, systems\, teams\, and incentive s to put AI at the heart of how we think\, build\, and deliver value. Thi s is a story in motion: a behind-the-scenes look at how we’re evolving fr om isolated AI feature experiments to an AI-native operating model.\n\nI’ ll share what we’re learning as we navigate the innovation dilemma—integr ating disruptive AI capabilities without breaking what already works or o ur roadmap. That includes rethinking how we define success\, how we hire\ , reward\, grow talent\, and how we handle legal and ethical complexity w ithout slowing down. We’ll explore the real-world tensions between rapid innovation\, value delivery\, making progress on Responsible AI\, all whi le elevating internal AI fluency\, and engaging with the broader AI ecosy stem to stay at the edge. \n\nThis isn’t a playbook from the finish line— it’s a candid reflection from deep inside the journey.\n\nMy goal is to h elp other leaders chart their own AI path with greater clarity\, confiden ce\, and care. DTEND:20250605T113500 DTSTAMP:20260403T174839Z DTSTART:20250605T111500 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:From Hype to Habit: How We’re Building an AI-First SaaS Company—Wh ile Still Shipping the Roadmap UID:SZSESSION914401 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Alvaro Morales\n\nAs AI continues to transform indust ries\, companies are faced with the critical challenge of effectively mon etizing AI-driven products in a way that captures value\, ensures custome r adoption\, and scales revenue sustainably. Unlike traditional SaaS mode ls\, AI-powered products have unique complexities - such as fluctuating u sage patterns\, variable compute costs\, and evolving customer demands\, making conventional pricing strategies unhelpful to the growth of an AI p roduct-led startup.\n\nIn this session\, Alvaro Morales\, CEO and co-foun der of Orb\, will explore why the often overlooked monetization aspect of AI is critical for businesses. He’ll share real-world examples and data to demonstrate how adaptive pricing models can drive cost savings\, enhan ce customer experience\, and reduce operational bottlenecks.\n\nAlvaro wi ll lead a live demo\, showcasing how engineers can simulate AI pricing st rategies and subsequently integrate them with a simple plug-and-play solu tion. He’ll also share how real-world revenue simulations enable companie s to test and refine pricing before implementing — reducing risk\, boosti ng adoption\, and unlocking new revenue streams. As a quick example\, clo ud software development platform Replit was looking to adopt a usage-base d pricing model for a new product\, but their existing billing system cou ldn't support the new model\, and building a new billing system would del ay the launch timeline. In order to get things done\, they turned to Orb\ , which enabled them to make pricing changes up to the last minute. After the launch\, Orb became the single source of truth for both Replit and i ts customers - providing usage alerts to notify Replit when users hit cos t thresholds and provide insights into user spend and payment methods.\n\ nKey takeaways: \nThe challenge of AI monetization – Why traditional subs cription-based SaaS pricing models don’t work for AI-powered products.\nP recision pricing – Exploring how usage-based\, tiered\, and hybrid pricin g models can maximize revenue potential. \nRevenue simulation for AI pric ing – Leveraging real-time data to test\, adjust and optimize pricing str ategies.\nAvoiding common pricing pitfalls – Identifying mistakes that ca n lead to revenue leakage and customer churn.\n\nThis session is designed for AI executives\, product leaders\, and engineering teams looking for actionable strategies to build adaptive\, scalable pricing models that dr ive long-term growth and profitability.\n\n DTEND:20250605T113500 DTSTAMP:20260403T174839Z DTSTART:20250605T111500 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Monetizing AI: From Zero to Profit UID:SZSESSION912033 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Scott Wu\n\nA talk on the future of software engineer ing with Scott Wu of Cognition AI\, the makers of Devin. DTEND:20250605T113500 DTSTAMP:20260403T174839Z DTSTART:20250605T111500 LOCATION:Yerba Buena Ballroom 7&8: SWE Agents SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Devin 2.0 and the Future of SWE UID:SZSESSION929855 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Will Brown\n\nThis talk will be a technical deep dive into RL for agentic reasoning via multi-turn tool calling\, similar to O penAI's o3 and Deep Research. In particular\, we'll cover:\n- When\, why\ , and how\n- GRPO vs PPO vs etc\n- Designing environments and rewards\n- Survey of recent research highlights\n- Results on example tasks\n- Overv iew of open-source ecosystem (libraries\, compute requirements\, tradeoff s\, etc.) DTEND:20250605T113500 DTSTAMP:20260403T174839Z DTSTART:20250605T111500 LOCATION:Yerba Buena Ballroom 2-6: Reasoning + RL SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Training Agentic Reasoners UID:SZSESSION914856 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jerry Liu\n\nAgents are all the rage in 2025\, and ev ery single b2b SaaS startup/incumbent promises AI agents that can "automa te work" in some way. \n\nBut how do you actually build this? The answer is two fold: \n1. really really good tools \n2. carefully tailored agent reasoning over these tools that range from assistant-to-automation based UXs. \n\nThe main goal of this talk is to a practical overview of agent architectures that can automate real-world work\, with a focus on documen t-centric tasks. Learn the core building blocks of best-in-class "tools" around processing\, manipulating\, and indexing/retrieving PDFs to Excel spreadsheets. Also learn the range of agent architectures suited for diff erent tasks\, from chat assistant-based UXs with high human-in-the-loop\, to automation UXs that rely on encoding a business process into an end-t o-end task solver. These architectures have to be generalizable but also highly accurate as agents get increasingly better at reasoning and code-w riting. DTEND:20250605T113500 DTSTAMP:20260403T174839Z DTSTART:20250605T111500 LOCATION:Golden Gate Ballroom A: Retrieval + Search SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building AI Agents that actually automate Knowledge Work UID:SZSESSION925259 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Omar Khattab\n\nWill discuss the principles for build ing AI software that underpin DSPy\, highlighting the differences between conventional prompting (or finetuning/RL) versus the design and programm ing of truly modular AI systems. DTEND:20250605T113500 DTSTAMP:20260403T174839Z DTSTART:20250605T111500 LOCATION:Golden Gate Ballroom B: Evals SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:On Engineering AI Systems that Endure The Bitter Lesson UID:SZSESSION936156 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Fouad Matin\n\nCode is the lingua franca for both sof tware engineers and highly capable AI models. As we give agents the abili ty to build\, test\, and run code that they generate\, the command line b ecomes their canvas—and their attack surface.\n\nThis keynote explores wh at it takes to bring code-executing agents from research to real-world de ployment while maintaining control and security. We’ll cover how terminal s offer AI an ideal interface\, why they’re deceptively risky\, and what it means to embed security\, guardrails\, and trust at every layer.\n\nIt ’s not just about what agents can do—it’s about what they should do\, and how we make sure they do it safely. DTEND:20250605T113500 DTSTAMP:20260403T174839Z DTSTART:20250605T111500 LOCATION:Foothill C: Security SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Safety and security for code-executing agents UID:SZSESSION938753 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Victor Dibia\n\nAutonomous or semi-autonomous multi-a gent systems (MAS) involve exponentially complex configurations (system c onfig\, agent configs\, task management and delegation\, etc.). These pre sent unique interface design challenges for both developer tooling and en d-user experiences.\nIn this session\, I'll explore UX design principles for multi-agent systems\, addressing critical questions: What is the true configuration space for autonomous MAS? How can users arrive at the corr ect mental model of an MAS's capabilities\, if at all? How can we improve trust and safety through techniques like cost-aware action delegation? W hat makes agent actions observable? How do we enable seamless interruptib ility? Attendees will gain actionable insights to create more transparent \, trustworthy\, and user-centered multi-agent applications\, illustrated through real-world implementations in AutoGen Studio - a low code develo per tool built on AutoGen (44k stars on GitHub\, MIT license) and similar tools. DTEND:20250605T113500 DTSTAMP:20260403T174839Z DTSTART:20250605T111500 LOCATION:Foothill G 1&2: Design Engineering SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:UX Design Principles for (Semi) Autonomous Multi-Agent Systems UID:SZSESSION914027 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Gorkem Yurtseven\n\nGenerative AI is reshaping the cr eative landscape\, enabling the production of images\, audio\, and video with unprecedented speed and sophistication. This session offers an in-de pth exploration of the current state of generative media\, highlighting c utting-edge models\, platforms\, and tools that are transforming the indu stry. DTEND:20250605T113500 DTSTAMP:20260403T174839Z DTSTART:20250605T111500 LOCATION:Foothill F: Generative Media SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The State of Generative Media Today UID:SZSESSION910158 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Jost Tobias Springenberg\, Quan Vuong\n\nSharing rec ent progress from Physical Intelligence and why it is an exciting time to push the frontier in general purpose robotics DTEND:20250605T113500 DTSTAMP:20260403T174839Z DTSTART:20250605T111500 LOCATION:Foothill E: Autonomy + Robotics SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Robotics: why now? UID:SZSESSION938258 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Adam Behrens\n\nHow to go beyond browser automation t o truly agentic commerce\, where AI can buy\, sell and negotiate on behal f of users and merchants. DTEND:20250605T115500 DTSTAMP:20260403T174839Z DTSTART:20250605T113500 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Machines of Buying & Selling Grace UID:SZSESSION914891 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Hariharan Ganesan\, Sahil Yadav\n\nEnterprise AI ado ption is accelerating\, but with it comes a hard question: Do we trust th e model’s decisions? In this 18-minute talk\, I’ll explore the invisible risks behind automated decision-making in safety-critical and revenue-sen sitive environments. Drawing on case studies across manufacturing\, telec om\, and industrial IoT\, I’ll highlight how explainability\, traceabilit y\, and robust guardrails drive adoption and protect enterprise value.\nA ttendees will walk away with:\n• A 3-step framework for operationalizing AI trust\n• Real-world lessons from building guardrails in on-prem and hy brid systems\n• Tools and techniques for debugging and explaining inferen ces at scale\n• A blueprint for building trust between models\, engineers \, and executive stakeholders DTEND:20250605T115500 DTSTAMP:20260403T174839Z DTSTART:20250605T113500 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:CIOs and Industry Leaders: Do You Trust Your AI’s Inferences? UID:SZSESSION916025 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Rustin Banks\n\nWill the future engineer code alongsi de a single coding agent\, or will they spend their day orchestrating man y agents? Traditional development rewards synchronous focus. This session dives into the significant mindshift required to move from sequential co ding to orchestrating parallel agents. We are the builders of "Jules"\, G oogle's massively parallel asynchronous coding agent (to be opened up in May). We'll share real-world insights from building Jules and explore how to rewire your brain for this powerful new "post-IDE" development paradi gm. DTEND:20250605T115500 DTSTAMP:20260403T174839Z DTSTART:20250605T113500 LOCATION:Yerba Buena Ballroom 7&8: SWE Agents SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Your Coding Agent Just Got Cloned And Your Brain Isn't Ready UID:SZSESSION914012 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Greg Kamradt\n\nARC Prize Foundation is building the North Star for AGI—rigorous\, open benchmarks that track reasoning progre ss in modern AI. We'll show why static AGI evaluations are useful\, but f all short when comparing models to human intelligence. Sneak peak preview of ARC-AGI-3: a dynamic\, game-like benchmark launching Q1 '26. DTEND:20250605T115500 DTSTAMP:20260403T174839Z DTSTART:20250605T113500 LOCATION:Yerba Buena Ballroom 2-6: Reasoning + RL SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Measuring AGI: Interactive Reasoning Benchmarks UID:SZSESSION914786 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Calvin Qi\, Chang She\n\nIn domains like law\, compl iance\, and tax\, building enterprise-grade RAG means very large scale\, spikey workloads\, a focus on accuracy\, and non-negotiable privacy.\nIn this talk\, we'll share war stories and battle scars of how Harvey has bu ilt the world's most advanced AI agents for the legal profession on top o f a highly optimized retrieval architecture. We'll cover how to get bette r retrieval via both sparse and dense retrieval methods\, why domain-spec ific reranking is essential\, and how to handle ambiguity in real-world q ueries.\nWe'll also touch on how LanceDB's search engine enables this arc hitecture by delivering low-latency\, high-throughput retrieval across mi llions of documents of varying sizes without compromising privacy. This s olid foundation enables Harvey to build a product that brings highly accu rate answers to hundreds of law firms and professional services firms acr oss 45 countries. DTEND:20250605T115500 DTSTAMP:20260403T174839Z DTSTART:20250605T113500 LOCATION:Golden Gate Ballroom A: Retrieval + Search SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Scaling Enterprise-Grade RAG Systems: Lessons from the Legal Front ier UID:SZSESSION903966 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Rafal Wilinski\, Vitor Balocco\n\nEvery agent failur e can be a roadmap to your next breakthrough. This talk reveals how Zapie r's evaluation system transforms frustrating user experiences into target ed improvements\, creating a data flywheel that continuously strengthens our agents. You'll learn practical approaches for building the data flywh eel\, detecting implicit feedback signals\, building solid evals\, priori tizing metrics that actually matter\, and why your most reliable evals mi ght secretly be sabotaging your performance. DTEND:20250605T115500 DTSTAMP:20260403T174839Z DTSTART:20250605T113500 LOCATION:Golden Gate Ballroom B: Evals SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Turning Fails into Features: Zapier’s Hard-Won Eval Lessons UID:SZSESSION936133 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jonathan Mortensen\n\nIn October 2024\, Apple release d a new private AI technology onto millions of devices called “Private Cl oud Compute”. It brings the same level of privacy and security a local de vice offers but on an “untrusted" remote server. This talk discusses how Private Cloud Compute represents a paradigm shift in confidential computi ng and explores the core advancements that made it possible to become mai nstream. We’ll explore its novel architecture that allows developers to r un sensitive\, multi-tenant workloads with cryptographically-provably pri vacy guarantees at scale and at reasonable cost. Attendees will leave wit h an understanding of how to leverage this technology for data and AI app lications where privacy and security is paramount. DTEND:20250605T115500 DTSTAMP:20260403T174839Z DTSTART:20250605T113500 LOCATION:Foothill C: Security SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The Unofficial Guide to Apple’s Private Cloud Compute UID:SZSESSION909905 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: John Pham\n\nBad designs are still bad. AI doesn’t ma ke it good. The novelty of AI makes the bad things tolerable\, for a shor t time. Building great designs and experiences with AI have the same firs t principles pre-AI. When people use software\, they want it to feel resp onsive\, safe\, accessible and delightful. We’ll go over the big and smal l details that goes into software that people want to use\, not forced to use. DTEND:20250605T115500 DTSTAMP:20260403T174839Z DTSTART:20250605T113500 LOCATION:Foothill G 1&2: Design Engineering SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Good design hasn’t changed with AI UID:SZSESSION914845 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Paige Bailey\n\nThis talk will briefly trace the hist ory of video generation models before diving into Veo 3\, Google DeepMind 's latest state-of-the-art model that marks a significant leap by generat ing video with synchronized audio—including dialogue\, sound effects\, an d music—all from text and image prompts. We'll show how it can understand ing intricate details\, maintain coherence over longer sequences\, and si mulate realistic physics and camera movements.\n\nFor developers\, Veo 3\ , accessible via Vertex AI (preview)\, unlocks many new capabilities. We' ll discuss how its advanced capabilities\, such as semantic context rende ring and cinematic control\, can empower innovation in filmmaking\, game development\, education\, and more. This session will cover how developer s can integrate Veo 3 into their workflows\, or test it out today in the Gemini App\, Flow\, and via the Gemini APIs on Google Cloud. DTEND:20250605T115500 DTSTAMP:20260403T174839Z DTSTART:20250605T113500 LOCATION:Foothill F: Generative Media SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Veo 3 for developers UID:SZSESSION943701 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Stefania Druga\n\nThe sheer volume of data and comple xity of modern scientific challenges necessitate tools that go beyond mer e analysis. The vision of an "AI Co-scientist" – a true collaborative par tner in the lab – requires sophisticated engineering to bridge the gap be tween powerful AI reasoning and the dynamic reality of physical experimen ts. This talk dives into the engineering required to build robust AI Co-s cientists for hands-on research. We will explore scalable architectures\, such as multi-agent systems leveraging foundation models like Gemini for complex reasoning\, hypothesis refinement (inspired by the "generate\, d ebate\, evolve" paradigm described in recent AI Co-scientist research)\, and intelligent tool use. The core focus will be on the engineering chall enges and solutions for integrating diverse\, real-time empirical data st reams – visual data from cameras\, quantitative readings from sensors\, p ositional feedback from actuators\, and instrument outputs – directly int o the AI's reasoning loop. I will illustrate this with concrete\, technic ally detailed examples in chemistry (adaptive reaction monitoring)\, robo tics (vision-guided assembly with SO Arm 100 and LeRobot library)\, and s ynthetic biology (real-time bacterial growth monitoring & interpretation) . We'll discuss engineering strategies for handling data heterogeneity\, latency\, noise\, and enabling the AI to interpret\, correlate\, and act upon live experimental feedback. Finally\, we will touch upon how thought ful engineering of these AI Co-scientists can contribute to democratizing access to advanced scientific capabilities.\n DTEND:20250605T115500 DTSTAMP:20260403T174839Z DTSTART:20250605T113500 LOCATION:Foothill E: Autonomy + Robotics SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Real-time Experiments with an AI Co-Scientist UID:SZSESSION916140 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jaspreet Singh\n\nI will talk about how Intuit uses L LMs to explain tax situations to Turbotax users.\nUsers want explanations of their tax situations - this drives confidence in the product. Over th e course of last two tax years\, Intuit has built out explanations using Anthropic and openAI’s models to develop genAI powered explanations. This includes design a complex system with prompt engineered solutions and bo th LLM & human powered evaluations to ensure high quality bar that our us ers expect when filing taxes with us.\nDuring the course of my talk\, I w ill talk across GenAI development lifecycle at scale - including developm ent \, evaluations and scaling. And security evaluations. We also develop ed a fine-tuned version of Claude Haiku & shall be covering that in the p resentation.\nWe also expanded into tax question and answering powered by RAG\, including graphRAG and I would be covering those developments too. DTEND:20250605T121500 DTSTAMP:20260403T174839Z DTSTART:20250605T115500 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How Intuit uses LLMs to explain taxes to millions of taxpayers UID:SZSESSION916085 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Yegor Denisov-Blanch\n\nForget vendor hype: Is AI act ually boosting developer productivity\, or just shifting bottlenecks? Sto p guessing.\n\nOur study at Stanford cuts through the noise\, analyzing r eal-world productivity data from nearly 100\,000 developers across hundre ds of companies. We reveal the hard numbers: while the average productivi ty boost is significant (~20%)\, the reality is complex – some teams even see productivity decrease with AI adoption.\n\nThe crucial insights lie in why this variance occurs. Discover which company types\, industries\, and tech stacks achieve dramatic gains versus minimal impact (or worse). Leave with the objective\, data-driven evidence needed to build a winning AI strategy tailored to your context\, not just follow the trend.\n DTEND:20250605T121500 DTSTAMP:20260403T174839Z DTSTART:20250605T115500 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Does AI Actually Boost Developer Productivity? (Stanford / 100k De vs Study) UID:SZSESSION914912 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Christopher Harrison DTEND:20250605T121500 DTSTAMP:20260403T174839Z DTSTART:20250605T115500 LOCATION:Yerba Buena Ballroom 7&8: SWE Agents SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The Agent Awakens: Collaborative Development with GitHub Copilot UID:SZSESSION936814 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Aakanksha Chowdhery\n\nThe models and techniques to b uild fully autonomous coding agents - not just coding copilots - are alre ady here. In this talk\, former Google DeepMind staff research scientist\ , now CEO of Reflection Misha Laskin will present new research on post-tr aining open weight LLMs for autonomous SWE tasks. He’ll focus on how scal ing LLMs with Reinforcement Learning improves the autonomous coding capab ilities of LLMs\, and provide insight on the technical challenges require d to train such systems at scale. DTEND:20250605T121500 DTSTAMP:20260403T174839Z DTSTART:20250605T115500 LOCATION:Yerba Buena Ballroom 2-6: Reasoning + RL SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Post-Training Open Models with RL for Autonomous Coding UID:SZSESSION915745 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Deanna Emery\, Julia Neagu\, Maitar Asher\n\nAI sear ch is becoming the front door to information\, whether through Retrieval- Augmented Generation (RAG)\, Search-Augmented Generation (SAG)\, or custo m agents that synthesize answers on top of indexed content. As users rely more heavily on these systems\, evaluating their quality becomes mission -critical. But traditional metrics like precision and recall don’t captur e the full picture.\n\nIn this talk\, we introduce a practical evaluation framework for AI-powered search\, across three dimensions:\n- Are the re trieved sources relevant to the query?\n- And is the final answer complet e?\n- Are the sources faithfully used in the generated answer?\n\nWe’ll s hare lessons from working with search companies and present early finding s from a new benchmark evaluating popular augmented AI systems across the se dimensions. Rather than ranking winners and losers\, we explore where different systems excel or break down\, and how these tradeoffs inform pr oduct decisions.\n\nThis talk is for AI engineers and product teams who w ant to build trusted\, high-quality AI search experiences\, and need a wa y to measure if it’s actually working. DTEND:20250605T121500 DTSTAMP:20260403T174839Z DTSTART:20250605T115500 LOCATION:Golden Gate Ballroom A: Retrieval + Search SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Evaluating AI Search: A Practical Framework for Augmented AI Syste ms UID:SZSESSION913839 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ido Pesok\n\nHow to think about evaluating a non-dete rministic system — and how to actually succeed at it. DTEND:20250605T121500 DTSTAMP:20260403T174839Z DTSTART:20250605T115500 LOCATION:Golden Gate Ballroom B: Evals SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Evals Are Not Unit Tests UID:SZSESSION939231 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Leonard Tang\n\n"Evaluation" is one of those concepts that every AI practitioner vaguely knows is important\, but few practiti oners truly understand. Is "eval" the dataset for measuring the quality o f your AI system? Is "eval" the measure\, the metric of quality? Is "eval " the process of human annotation and scoring? Or is "eval" a third-party dataset run once to benchmark a model?\n\nTo mitigate this cacophony\, t his talk will provide an opinionated and principled perspective for what we actually mean when we say “evaluation”\, beyond the traditional for-lo op-over-a-static dataset. \n\nIn particular\, this perspective draws heav y inspiration from *fuzzing*\, i.e. bombarding AI with simulated\, unexpe cted user inputs to uncover corner cases at scale. This factors into sub- problems regarding:\n\n- Quality Metric. What is the actual criteria we\, as humans\, are using to determine if an AI system is producing good or bad responses? How do we elicit these criteria before the human SME can a rticulate them? How do we\, as efficiently as possible\, operationalize t his criteria with an automated *Judge*?\n\n- Stimuli Generation. Given a metric\, how do we know\, with confidence\, that an AI system is performi ng well with respect to the metric? What data is representative and suffi cient for discovering all potential bugs of an AI system? And how do we g enerate this complex\, diverse\, faithful data at scale? \n\nWe will disc uss in detail the philosophy\, technology\, and case studies behind both problems of Quality Metric and Stimuli Generation\, and how they interact in concert. DTEND:20250605T121500 DTSTAMP:20260403T174839Z DTSTART:20250605T115500 LOCATION:Foothill C: Security SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Fuzzing in the GenAI Era UID:SZSESSION915978 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Maximillian Piras\n\nAre conversational interfaces th e future or\, as many designers have suggested\, a lazy solution that is bottlenecking AI-HCI? Despite well-documented usability issues\, the desi gn of many AI applications defaults to an input field\, turn-by-turn flow \, and an endless model picker — I call this “The Bitter Layout”. \n\nIn this talk\, we’ll explore how Clay Christensen’s theory of commoditizatio n from the early PC industry can explain why scaling laws require AI inte rfaces to remain modular until models fully commoditize. The killer featu re of conversational interfaces may not be that they’re natural\, but tha t they’re conformable. Learn how to evolve interfaces as inference scales \, spot shifts in the basis of competition\, and stop worrying about the next model update steamrolling your design decisions. DTEND:20250605T121500 DTSTAMP:20260403T174839Z DTSTART:20250605T115500 LOCATION:Foothill G 1&2: Design Engineering SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The Bitter Layout or: How I Learned to Love the Model Picker UID:SZSESSION915428 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Kelvin Ma\n\nGo behind the scenes of Google Photos' M agic Editor. Explore the engineering feats required to integrate complex CV and cutting-edge generative AI models into a seamless mobile experienc e. We'll discuss optimizing massive models for latency/size\, the crucial interplay with graphics rendering (OpenGL/Halide)\, and the practicaliti es of turning research concepts into polished features people actually us e. DTEND:20250605T121500 DTSTAMP:20260403T174839Z DTSTART:20250605T115500 LOCATION:Foothill F: Generative Media SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Magic Editor Under the Hood: Weaving Generative AI into a Billion- User App UID:SZSESSION910197 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Aastha Jhunjhunwala\, Annika Brundyn\n\nFoundation m odels don’t just write or draw anymore—they’re starting to move.\n\nGR00T N1 is NVIDIA’s open Vision-Language-Action (VLA) foundation model for hu manoid robots. Built with a dual-system architecture\, it combines a Syst em 2 module for high-level reasoning with a System 1 module for real-time \, fluid motor control. It’s trained end-to-end on a an impressive mix of data—from human videos to robot trajectories to synthetic simulations—an d deployed on a full-sized humanoid robot performing bimanual manipulatio n tasks in the real world.\nThis talk is a high-level\, beginner-friendly overview of GR00T N1:\n- What makes a robot foundation model different f rom an LLM or vision model\n- How GR00T’s architecture is inspired by cog nitive systems\n- Why grounding language\, vision\, and action together u nlocks new generalist capabilities\n\nIf you’ve ever wondered how large-s cale AI is crossing over into the physical world\, this session will get you up to speed—no robotics PhD required. DTEND:20250605T121500 DTSTAMP:20260403T174839Z DTSTART:20250605T115500 LOCATION:Foothill E: Autonomy + Robotics SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:What Is a Humanoid Foundation Model? An Introduction to GR00T N1 UID:SZSESSION916103 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Mike Chambers\n\nExplore the practical challenges and solutions for deploying AI agents in real-world production environments. Through detailed technical analysis and practical examples\, we'll exami ne strategies for building and orchestrating agent systems at scale. We'l l cover critical infrastructure decisions\, scalability frameworks\, and best practices for creating robust\, production-ready agent architectures . DTEND:20250605T123500 DTSTAMP:20260403T174839Z DTSTART:20250605T121500 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Ship it! Building Production-Ready Agents UID:SZSESSION933610 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ivan Burazin\n\nIf you’re building devtools for human s\, you’re building for the past. \n\nAlready a quarter of Y Combinator’s latest batch used AI to write 95% or more of their code. AI agents are s caling at an exponential rate and soon\, they’ll outnumber human develope rs by orders of magnitude.\n\n\nThe real bottleneck isn’t intelligence. I t’s tooling. Terminals\, local machines\, and dashboards weren’t built fo r agents. They make do… until they can’t.\n\nIn this talk\, I’ll share ho w we killed the CLI at Daytona\, rebuilt our infrastructure from first pr inciples\, and what it takes to build devtools that agents can actually u se. Because in an agent-native future\, if agents can’t use your tool\, n o one will.\n DTEND:20250605T123500 DTSTAMP:20260403T174839Z DTSTART:20250605T121500 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:AX is the only Experience that Matters UID:SZSESSION914814 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Tomas Reimers\n\nAs AI tools like GitHub Copilot and ChatGPT help engineers generate code at an unprecedented rate\, the “oute r loop”—reviewing\, testing\, merging\, and deploying—becomes more vital than ever. Studies have shown that up to half of AI-generated solutions c ontain bugs or vulnerabilities\, underscoring the continued importance of thorough\, human-in-the-loop reviews. In this talk we'll take a look at how next-gen developer tools can harness AI not just for generating code\ , but also reviewing it. By thoughtfully integrating AI into that fully u nderstands your entire codebase\, teams can accelerate velocity without s acrificing quality.\n\nAttendees will learn real-world strategies and bes t practices for establishing an “outer loop” that safely and efficiently deploys high volumes of AI-assisted code\, without compromising reliabil ity. We’ll also discuss pitfalls to avoid when integrating AI into existi ng pipelines. DTEND:20250605T123500 DTSTAMP:20260403T174839Z DTSTART:20250605T121500 LOCATION:Yerba Buena Ballroom 7&8: SWE Agents SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Don’t get one-shotted: Leveraging AI to test\, review\, merge\, an d deploy code UID:SZSESSION933458 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ryan Marten\n\nPeel back the curtain on state of the art model post-training through the story of OpenThinker\, a SOTA small r easoning model (outperforming DeepSeek distill)\, built in the open. Lear n about the dataset recipe used to build the strongest reasoning models w hich you can apply to your own domain-specific specialized reasoning mode ls. Hear about the strategies that scale (and that don't) based on our ri gorous experimentation on the journey from thousands of data points (Besp oke-Stratos) to millions of data (OpenThinker3). Build upon our open sour ce engineering solutions for large-scale synthetic data generation\, trai ning on multiple supercomputing clusters\, and building out fast reliable evaluations. DTEND:20250605T123500 DTSTAMP:20260403T174839Z DTSTART:20250605T121500 LOCATION:Yerba Buena Ballroom 2-6: Reasoning + RL SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:OpenThoughts: Data Recipes for Reasoning Models UID:SZSESSION916074 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Tengyu Ma\n\n The talk will have three parts\n1.Roadm ap debate: RAG vs. finetuning vs. long-context\n2.RAG today: benefits\, c hallenges\, and current solutions\n3.RAG tomorrow: AI models do more work DTEND:20250605T123500 DTSTAMP:20260403T174839Z DTSTART:20250605T121500 LOCATION:Golden Gate Ballroom A: Retrieval + Search SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:RAG in 2025: State of the Art and the Road Forward UID:SZSESSION933678 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Diego Rodriguez\n\nSpecial session with KREA.ai's cof ounder Diego Rodriguez on how evals for aesthetics and image/generative m edia work — the hardest kinds of evals. DTEND:20250605T123500 DTSTAMP:20260403T174839Z DTSTART:20250605T121500 LOCATION:Golden Gate Ballroom B: Evals SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Perceptual Evaluation UID:SZSESSION949432 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Bobby Tiernay\, Kam Sween\n\nShipping AI agents that are safe for production means solving some tough identity and authorizat ion challenges that are not always obvious at the prototype stage. In pra ctice\, this comes down to a handful of deeply technical questions:\n- Ho w do you make sure agents are only acting for the right user?\n- How do y ou prevent over-broad API access or data leaks?\n- How do you handle user approvals when there is no UI\, or you need a human in the loop?\n- And how do you avoid the usual pain points like manual credential sharing\, s tale keys\, or unpredictable scopes without writing a lot of brittle\, cu stom code?\n\nThis talk digs into the real technical trade-offs behind bu ilding secure\, user-aware AI agents. We will go beyond what to do and ex plain why\, sharing the architectural decisions\, open standards\, and ha rd lessons learned from integrating OAuth\, OIDC\, RAR\, and async author ization into agent-driven workflows.\n\nYou will see a hands-on demo usin g an open-source Node.js agent and open protocols\, with a focus on pract ical integration and no magic. The session will show how these solutions have shaped our approach to identity in GenAI and where we see the field heading next.\n\nIf you are an engineer building AI apps that need real g uardrails\, not just a happy-path demo\, we hope to leave you with some p ractical patterns\, design rationale\, and a clear view of the trade-offs for making your own agents production ready. DTEND:20250605T123500 DTSTAMP:20260403T174839Z DTSTART:20250605T121500 LOCATION:Foothill C: Security SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Securing Agents with Open Standards UID:SZSESSION936795 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Shafik Quoraishee\n\nThis session will examine the in terplay between human intuition and artificial intelligence in puzzle-sol ving\, using the popular New York Times Connections game as a practical c ase study. \n\nWe'll investigate how gameplay can be systematically evalu ated through AI algorithms\, exploring machine learning strategies such a s clustering\, semantic mapping\, and natural language processing. \n\nAt tendees will gain insights into building AI-driven puzzle solvers\, learn methods for quantitatively assessing gameplay complexity\, and discuss t he potential impacts of AI on puzzle game design and player engagement. DTEND:20250605T123500 DTSTAMP:20260403T174839Z DTSTART:20250605T121500 LOCATION:Foothill G 1&2: Design Engineering SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:AI and Game Theory: A Case Study on NYT's Connections UID:SZSESSION914361 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Zeke Sikelianos\n\nLegendary AI engineer and educator Andrej Karpathy recently blogged about his experiences building\, deploy ing\, and monetizing a vibe-coded web app called MenuGen. Let's dig into the challenges he faced and learn what we as AI designers can do to make life better for the Andrejs of the world. DTEND:20250605T123500 DTSTAMP:20260403T174839Z DTSTART:20250605T121500 LOCATION:Foothill F: Generative Media SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Design like Karpathy is watching 😎 UID:SZSESSION943296 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Rishabh Garg\n\nA journey into building a small softw are stack for a robot and discussing the issues that may commonly come up along the way. DTEND:20250605T123500 DTSTAMP:20260403T174839Z DTSTART:20250605T121500 LOCATION:Foothill E: Autonomy + Robotics SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Communication and System Software in Robotics UID:SZSESSION949382 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250605T133000 DTSTAMP:20260403T174839Z DTSTART:20250605T123000 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Lunch UID:SZSESSION927cb841-b6c9-42b3-b002-5128c980861e END:VEVENT BEGIN:VEVENT DESCRIPTION:[last round of Attendee-Led 10min lightning talks]\n\n\nPracti cal tactics to build reliable AI apps. Reverse engineering real-world eva ls with o3. Nobody does it this way.\nCompanies pay me $500/h for this kn owledge. I help them get from POC that works 50% of the time - to the sol ution they can trust to deploy to production. DTEND:20250605T125500 DTSTAMP:20260403T174839Z DTSTART:20250605T124500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Practical tactics to build reliable AI apps — Dmitry Kuchin UID:SZSESSION870f5471-3b04-47ce-b0c7-c4c8ef5d88ea END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Den Delimarsky (DEVDIV)\, Julia Kasper\n\nJoin us to see how VS Code and GitHub Copilot's expanding suite of AI features can match or even surpasses the benefits of other popular AI developer tools. We'll focus on practical scenarios to ensure immediate applicability an d work through live demos of Copilot features such as: Code generation us ing Edits\, Planning/problem solving using Chat\, Inline terminal command generation\, Boilerplate code generation using Agent mode\, Improving bo ilerplate with custom instructions and then refactoring using Agent mode and Edits\, Improving test generation and code reviews with custom instru ctions\, as well as an Introduction to MCP. \n DTEND:20250605T130500 DTSTAMP:20260403T174839Z DTSTART:20250605T124500 LOCATION:Nobhill C&D: Microsoft SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Protected MCP Servers UID:SZSESSION936816 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Zach Blumenfeld\n\nAgentic workflows often become com plex\, brittle\, and hard to maintain when they need to retrieve and reas on across both structured data (typically requiring precise query executi on) and unstructured data (commonly handled via vector search in RAG). In this talk\, we’ll explore how mapping key information into a knowledge g raph can simplify these workflows and improve retrieval quality. You’ll l earn core concepts behind GraphRAG\, how to integrate it into agent tools \, and get access to end-to-end code examples so you can start building r ight away. DTEND:20250605T130000 DTSTAMP:20260403T174839Z DTSTART:20250605T124500 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Agentic GraphRAG: Simplifying Retrieval Across Structured & Unstru ctured Data UID:SZSESSION933671 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Nick Nisi\n\nAI agents are calling APIs\, submitting forms\, and sending emails—but how do you control what they’re allowed to do? As agents act on behalf of users or organizations\, traditional patt erns like OAuth\, session tokens\, and role-based access often fall short .\nIn this talk\, we’ll explore how machine identity is evolving to meet this new landscape. You’ll learn:\n\n- How to think about authentication for agents (not just humans)\n- What it means to authorize an action when the actor is an LLM or headless service\n- Real-world strategies from Wo rkOS and Cloudflare for assigning\, managing\, and revoking agent identit y and access\n\nBy the end\, you’ll walk away with practical tools and me ntal models to build agent-powered systems that are secure\, auditable\, and scalable. DTEND:20250605T130000 DTSTAMP:20260403T174839Z DTSTART:20250605T124500 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Agents\, Access\, and the Future of Machine Identity UID:SZSESSION933656 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Lou Bichard\n\nSecurity is the biggest blocker for ag ent orchestration adoption in regulated industries for SWE agents. Gitpod 's agent orchestration went from an originally self-hosted kubernetes arc hitecture to the current 'bring your own cloud' model that enables deploy ment our SWE agent orchestration platform in secure environments. The arc hitecture allows customers to securely connect their foundational models and agent memory solutions and comes with features like auto-suspend and resume for agent fleets. In this talk we deep dive into the architecture to share our years of learnings in how to secure agent workloads at scale in secure and regulated environments. DTEND:20250605T130000 DTSTAMP:20260403T174839Z DTSTART:20250605T124500 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building CISO-approved agent fleet architecture UID:SZSESSION933569 END:VEVENT BEGIN:VEVENT DESCRIPTION:[last round of Attendee-Led 10min lightning talks]\n\nHow to r un evals at scale? Thinking beyond accuracy or similarity.\nMuktesh Mishr a DTEND:20250605T130500 DTSTAMP:20260403T174839Z DTSTART:20250605T125500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How to run evals at scale? Thinking beyond accuracy or similarity — Muktesh Mishra UID:SZSESSION139b04f9-d334-4fa6-8195-811051b679cf END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Philipp Krenn\n\nMCP is a solid integration layer — b ut how does it hold up when it comes to output quality? Often\, not as we ll as you'd like. Here are some practical retrieval patterns\, from basic to advanced\, that worked well in my experiments:\n* Naive: Just plug in plain MCP and hope the LLM gets it right. Sometimes it does. Sometimes y ou’ll need a miracle.\n* Semantic: Add more descriptive field names and e xtra metadata. It helps — but usually just a bit.\n* Templated: Use a str uctured template and have the LLM fill it out step by step. More effort\, but by far the most reliable results. DTEND:20250605T131500 DTSTAMP:20260403T174839Z DTSTART:20250605T130000 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Hope is Not a Strategy: Retrieval Patterns for MCP UID:SZSESSION933626 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Doug Guthrie\n\nThis hands-on workshop guides partici pants through the full AI evaluation lifecycle with Braintrust\, from ini tial prompt testing to production monitoring. Attendees will build evalua tion frameworks\, practice offline and online strategies\, and implement logging systems. DTEND:20250605T134500 DTSTAMP:20260403T174839Z DTSTART:20250605T130000 LOCATION:Golden Gate Ballroom B: Evals SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Evals 101: Lunch & Learn UID:SZSESSION949122 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Samuel Colvin\n\nIn this talk I'll introduce the conc ept of Human-seeded Evals\, explain the principle and demo them with Pyda ntic Logfire. DTEND:20250605T131500 DTSTAMP:20260403T174839Z DTSTART:20250605T130000 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Human-seeded Evals UID:SZSESSION933676 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Rick Blalock\n\nAgent deployments can be dicey\, espe cially at first. This session goes over all the things that cause headac he with deployments from serverless issues to networking issues - and how we fix them. DTEND:20250605T131500 DTSTAMP:20260403T174839Z DTSTART:20250605T130000 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Conquering Agent Chaos UID:SZSESSION933618 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Harald Kirschner\n\n"Vibe coding" often falters in co mplex enterprise environments. Drawing from real implementations\, this t alk demonstrates systematic approaches to customizing AI assistants for c hallenging codebases. We'll explore specialized techniques for navigating complex architectures\, evidence-based strategies for undocumented legac y systems\, methodologies for maintaining context across polyglot environ ments\, and frameworks for standardizing AI usage while preserving develo per autonomy. Through case studies from finance and healthcare\, we'll pr esent a comprehensive evaluation framework that bridges the gap between A I's theoretical capabilities and practical enterprise implementation\, en abling true flow-state collaboration even within the most complex develop ment ecosystems. DTEND:20250605T131500 DTSTAMP:20260403T174839Z DTSTART:20250605T130000 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Vibe Coding at Scale: Customizing AI Assistants for Enterprise Env ironments UID:SZSESSION936902 END:VEVENT BEGIN:VEVENT DESCRIPTION:[last round of Attendee-Led 10min lightning talks]\n\nHey All - gave a talk on building stateful environments for vertical agents at AI tinkerers and ppl really liked it\, happy to do again. Here's the repo - general code that endows environments like Pokemon Red\, Minecraft\, Swe -Bench\, and others with the same interface for development and agent tra ining. github.com/synth-laboratories/Environments DTEND:20250605T131500 DTSTAMP:20260403T174839Z DTSTART:20250605T130500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Stateful environments for vertical agents — Josh Purtell UID:SZSESSION453b3f4c-4d50-4c33-8fb9-8b3ed9ee1831 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jon Peck\n\n"Software development is a team sport\, w ith many different roles\, where eveyone can win. But success isn't guara nteed\; it depends on specific practices\, policies\, and tools which ena ble minimally-siloed\, AI-accelerated collaboration across all parts of t he DevOps process\, from PM to development to CI/CD and security.\n \n Di scover the patterns and tools which lead to success\, methods for changin g the status quo\, and perhaps a few horror stories. We'll touch on inner sourcing\, cloud development\, AI\, automation\, governance\, security\, scaling and more -- with actionable learnings for everyone from small mai ntainer communities to F500 Enterprises." DTEND:20250605T133000 DTSTAMP:20260403T174839Z DTSTART:20250605T131000 LOCATION:Nobhill C&D: Microsoft SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Unlocking AI-Powered DevOps Within Your Organization UID:SZSESSION936901 END:VEVENT BEGIN:VEVENT DESCRIPTION:[last round of Attendee-Led 10min lightning talks]\n\nAre your vibes immaculate? - Vibe coding is the new hotness but everyone has a st ory of AI making really dumb choices. Let's talk about how you can improv e your vibe coding so your vibes are safe and bug free and you spend more \nIan Butler DTEND:20250605T132500 DTSTAMP:20260403T174839Z DTSTART:20250605T131500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Are your vibes immaculate? Improve your Vibe Coding — Ian Butler UID:SZSESSIONf8a76c2c-fc05-4fb0-83fd-de0ac064d873 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Mikiko Bazeley\n\nIn today’s most advanced AI systems \, intelligence is no longer confined to a single model or agent—it emerg es from coordination. But coordination requires memory: short-term\, long -term\, and shared. In this talk\, we’ll break down how agent systems can store\, retrieve\, and evolve shared memory to become smarter over time. You'll learn what it takes to architect these continuously learning syst ems\, how to track and improve memory quality\, and why robust\, flexible infrastructure is the foundation of it all. Stick around to see how this works in practice—live. DTEND:20250605T133000 DTSTAMP:20260403T174839Z DTSTART:20250605T131500 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Smarter Together: Designing Multi-Agent Systems with Shared\, Evol ving Memory UID:SZSESSION933702 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Arjun Desai\, Rohit Talluri\n\nReal-Time Voice AI a pplications demand the lowest possible latencies to enhance user experie nces with more advanced reasoning and agentic capabilities. AWS is hosti ng Arjun Desai\, co-founder of Cartesia\, in a fireside chat for a techn ical deep dive into learnings and best practices for building a state-of -the-art inference stack that serves global enterprise customers. DTEND:20250605T133000 DTSTAMP:20260403T174839Z DTSTART:20250605T131500 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Serving Voice AI at Scale UID:SZSESSION933599 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Philip Kiely\n\nHow do you get time to first byte (TT FB) below 150 milliseconds for voice models -- and scale it in production ? As it turns out\, open-source TTS models like Orpheus have an LLM backb one that lets us use familiar tools and optimizations like TensorRT-LLM a nd FP8 quantization to serve the models with low latency. But client code \, network infrastructure\, and other outside-the-GPU factors can introdu ce latency in the production stack. In this talk\, we'll cover the basic mechanics of TTS inference\, common pitfalls to avoid in integrating them into production systems\, and how to extend this high-performance system to serve customized models with voice cloning and fine-tuning. DTEND:20250605T133000 DTSTAMP:20260403T174839Z DTSTART:20250605T131500 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Optimizing inference for voice models in production UID:SZSESSION933580 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Nir Gazit\n\nManual prompt crafting doesn't scale. In this session\, we'll explore how to replace it with a test-driven\, auto mated approach. You'll see how to define output evaluators\, write minima l prompts\, and let agents iterate toward optimal performance—all without manual tweaking. If you're still hand-tuning prompts\, you're doing it w rong. DTEND:20250605T133000 DTSTAMP:20260403T174839Z DTSTART:20250605T131500 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Prompt Engineering is Dead UID:SZSESSION933712 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Hf0 Residency\n\n10 CEOs. 2 minutes each. Cutting edg e voice models. Post-transformer architectures. Game changing tech. Insan e business plans. Be the first to hear updates from the leaders of 10 of the fastest growing Seed+ & Series A startups in the world. Who do you th ink will be the first to hit $1B?\n\n\nTeams:\nArea\nOpenRouter\nFavorite d\nOpenAudio\nCoframe\nOpenHome\nUpside\nRecursal\nGlow\nGeneration Lab DTEND:20250605T140000 DTSTAMP:20260403T174839Z DTSTART:20250605T132500 LOCATION:Golden Gate Ballroom A: Retrieval + Search SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:[New Session] The Next AI Unicorns UID:SZSESSION949405 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Chad Bailey\n\nLearn about building voice agents with for customer support\, call center workflows\, market research\, and man y other use cases. Pipecat is the open source\, vendor neutral realtime agent framework used by teams at NVIDIA\, OpenAI\, Google DeepMind\, AWS\ , and hundreds of startups and scale-ups. DTEND:20250605T134500 DTSTAMP:20260403T174839Z DTSTART:20250605T133000 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Pipecat Cloud: Open Source Enterprise Voice AI UID:SZSESSION933491 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Tomas Reimers\n\nIt’s not enough that AI can review a nd understand your code in a vacuum. In order to be effective\, AI agents need context around your entire codebase and understand the nuances of t he way your team works. This talk will examine technical approaches and b est practices for teaching AI tools to enforce highly customized style gu ides and coding conventions no matter what tech stack you’re working in. DTEND:20250605T140000 DTSTAMP:20260403T174839Z DTSTART:20250605T134500 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Onboarding your robot coworkers: customizing AI code review agents to your team’s style UID:SZSESSION933453 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Yogi Miraje\n\nLLMs are getting smarter—but Agents ar e still unpredictable\, unreliable\, and hard to control.\n\nIn this talk \, I’ll share practical lessons from building real-world plan-and-execute agents —covering how to steer autonomous agents using agentic workflows\ , blueprints\, and evals.\n\nIf you’re struggling to make your agents beh ave (without giving up flexibility)\, this one’s for you.\n DTEND:20250605T142000 DTSTAMP:20260403T174839Z DTSTART:20250605T140000 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How to Build Agents without losing control UID:SZSESSION914049 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Michael Grinich\n\nAI agents are changing the way mod ern SaaS products operate. Whether automating workflows\, integrating wit h APIs\, or acting on behalf of users\, AI-driven assistants and autonomo us systems are becoming core product features. But securing these agents presents a fundamental challenge: How do you authenticate AI agents? How do you control what they can access? How do you ensure they act within th e right permissions? This talk will explore these concepts and more while highlighting current research and best practices. DTEND:20250605T142000 DTSTAMP:20260403T174839Z DTSTART:20250605T140000 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:CIAM for AI: Who Are Your Agents and What Can They Do? UID:SZSESSION933686 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Boris Cherny\n\nA ten thousand foot view of the codin g space\, the UX of coding\, and the Claude Code team's approach DTEND:20250605T142000 DTSTAMP:20260403T174839Z DTSTART:20250605T140000 LOCATION:Yerba Buena Ballroom 7&8: SWE Agents SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Claude Code & the evolution of agentic coding UID:SZSESSION939942 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Kyle Corbitt\n\nHave you ever launched an awesome age ntic demo\, only to realize no amount of prompting will make it reliable enough to deploy in production? Agent reliability is a famously difficult problem to solve!\n\nIn this talk we’ll learn how to use GRPO to help yo ur agent learn from its successes and failures and improve over time. We’ ve seen dramatic results with this technique\, such as an email assistant agent that whose success rate jumped from 74% to 94% after replacing o4- mini with an open source model optimized using GRPO.\n\nWe’ll share case studies as well as practical lessons learned around the types of problems this works well for and the unexpected pitfalls to avoid. DTEND:20250605T142000 DTSTAMP:20260403T174839Z DTSTART:20250605T140000 LOCATION:Yerba Buena Ballroom 2-6: Reasoning + RL SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How to Train Your Agent: Building Reliable Agents with RL UID:SZSESSION914533 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Satwik Singh\, Sherwood Callaway\n\nAI agents are be coming essential tools for teams of all sizes and industries - but traini ng them to become experts in your product\, business\, and customerbase r emains a challenge. \n\nWhat if onboarding a digital worker was as simple as uploading your pitch deck? At 11x\, we built Alice\, an AI SDR that w rites outbound emails with the nuance and context of a top-performing hum an sales rep - because she learns like one too!\n\nIn this talk\, we'll s hare how we built a knowledge base that allows 11x customers to "train" A lice on their internal materials: PDFs\, websites\, call recordings\, and more. We'll talk through the ingestion pipeline in detail\, discuss stor age/retrieval technologies and their tradeoffs\, and explain how Alice us es the knowledge base to drive high-performance email outreach at scale. DTEND:20250605T142000 DTSTAMP:20260403T174839Z DTSTART:20250605T140000 LOCATION:Golden Gate Ballroom A: Retrieval + Search SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Alice’s Brain: How We Built an AI Sales Rep that Learns L ike a Human UID:SZSESSION916215 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ankur Goyal\n\ntbc DTEND:20250605T142000 DTSTAMP:20260403T174839Z DTSTART:20250605T140000 LOCATION:Golden Gate Ballroom B: Evals SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:[Evals Keynote] tba UID:SZSESSION939130 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: David Mytton\n\nConstantly seeing CAPTCHAs? It used t o be easy to detect the humans from the droids\, but what else can we do when synthetic clients make up nearly half of all web requests. Rotating IPs\, spoofed browsers\, and agents acting on behalf of real users - are we doomed to forever be solving puzzles?\n\nIn this talk\, we’ll explore user agents\, HTTP fingerprints\, and IP reputation signals that make hum ans and agents stand out from scrapers\, build a realistic threat model\, and dig into the behaviors that reveal the LLM-mimicry. Leave with AX- a nd UX-safe code\, benchmarks\, and tools to help you take back control. DTEND:20250605T142000 DTSTAMP:20260403T174839Z DTSTART:20250605T140000 LOCATION:Foothill C: Security SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How to defend your sites from AI bots UID:SZSESSION913351 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Christopher Chedeau\n\nCovid sent everybody home and created the space of virtual whiteboards. At first the experience reused the physical constraints but soon it became better than a physical whiteb oard thanks to using virtual native concepts like copy-paste and using ke yboard input.\nThe next step in this evolution is to integrate AI into th e workflow. We've tried a lot of things with Excalidraw and ended up land ing on turning prompt into diagram. Come to the talk to understand how it fits into the workflow and how we implemented it. DTEND:20250605T142000 DTSTAMP:20260403T174839Z DTSTART:20250605T140000 LOCATION:Foothill G 1&2: Design Engineering SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:AI and Human Whiteboarding Partnership UID:SZSESSION915389 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Keegan McCallum\n\nTalking about Luma AI\, our missio n\, and how our ML infrastructure enables SOTA multimodal model developme nt DTEND:20250605T142000 DTSTAMP:20260403T174839Z DTSTART:20250605T140000 LOCATION:Foothill F: Generative Media SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:General Intelligence is Multimodal UID:SZSESSION914081 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jyh-Jing Hwang\n\nThis session explores Waymo's lates t research on the End-to-End Multimodal Model for Autonomous Driving (EMM A) and advanced sensor simulation techniques. Jyh-Jing Hwang will demonst rate how multimodal large language models like Gemini could improve auton omous driving through unified end-to-end architectures that process raw s ensor data directly into driving decisions. \n\nThe presentation will sho wcase EMMA's state-of-the-art performance in trajectory planning\, 3D obj ect detection\, and road graph understanding\, as well as another Drive&G en research approach to sensor simulation for evaluating an end-to-end mo tion planning model. Attendees will gain insights into the benefits of co -training across multiple autonomous driving tasks and the potential of c ontrolled video generation for testing under various environmental condit ions.\n\nMore on EMMA here: https://waymo.com/blog/2024/10/introducing-em ma\n DTEND:20250605T142000 DTSTAMP:20260403T174839Z DTSTART:20250605T140000 LOCATION:Foothill E: Autonomy + Robotics SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Teaching Cars to Think: Language Models and Autonomous Vehicles UID:SZSESSION915934 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Randall Hunt\n\nThe transition from experimental GenA I demonstrations to robust\, production-grade systems involves significan t technical and organizational complexities. Humans provide a ceiling on the true ROI of automations. This session synthesizes key patterns and pr actical strategies gathered from more than 200 GenAI implementations acro ss multiple industries and business sizes.\n\nBeyond the general lessons that apply to most products leveraging GenAI\, we'll cover detailed obser vations within three application areas: multimodal understanding and sear ch\, enterprise knowledge retrieval\, and AI agent architectures. We will share real-world comparative performance data and metrics on embedding m odels\, vector index implementations\, and explore various implementation methodologies that balance performance and cost.\n\nAdditionally\, the s ession addresses organizational insights critical to successful AI deploy ments\, such as the importance of clearly defined evaluation processes an d understanding real-world user interaction challenges\, highlighted by e xamples from healthcare environments. Attendees will gain an understandin g of decision-making criteria\, including the appropriate complexity of p rompt engineering versus more elaborate orchestration methods\, token/cos t management strategies in multilingual settings\, and the challenges in driving behavioral change with new UX and application interaction capabil ities.\n\nParticipants will leave equipped with practical\, data-supporte d insights for effectively navigating their own GenAI projects\, includin g benchmarks and criteria for informed technology selection\, and techniq ues to streamline the transition from initial concept to sustainable oper ational deployment. Please note\, we all know this field evolves rapidly and we will mark which lessons we believe are immutable. DTEND:20250605T144000 DTSTAMP:20260403T174839Z DTSTART:20250605T142000 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:POC to PROD: Hard Lessons from 200+ Enterprise GenAI Deployments UID:SZSESSION925912 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Ben Hylak\, Sid Bendre\n\nYou've made the demo. How do you make the product? A lot of AI products don't actually work. Even w orse\, a lot of the techniques being advertised for making AI products be tter don't work either. We'll cover the challenges + techniques we've see n actually work in the real world. DTEND:20250605T144000 DTSTAMP:20260403T174839Z DTSTART:20250605T142000 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building AI Products That Actually Work UID:SZSESSION915921 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Robert Brennan\n\nThe adoption of AI into software de velopment has been bumpy. While autocomplete tools like Copilot have gone mainstream\, autonomous agents like Devin and OpenHands have generated b oth enthusiasm and skepticism. Some engineers claim they generate a 10x p roductivity boost\; others that they just create noise and tech debt.\n\n The difference between the enthusiasts and the skeptics is that the enthu siasts have reasonable expectations for what these agents can do\, and ha ve both practical and intuitive knowledge for how to use them effectively .\n\nIn this session\, we'll talk about what tasks are appropriate for to day's software agents\, what tasks they might start to succeed at in 2025 \, and what tasks are best left to humans no matter how good they get.\n\ nSession Outline:\nLearn how to use software development agents like Open Hands (fka OpenDevin) effectively\, without creating noise and tech debt. DTEND:20250605T144000 DTSTAMP:20260403T174839Z DTSTART:20250605T142000 LOCATION:Yerba Buena Ballroom 7&8: SWE Agents SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Software Development Agents: What Works and What Doesn't UID:SZSESSION915387 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Nathan Lambert\n\nCurrent AI models are extremely ski lled\, which was seen as the step change in evaluation scores across the industry in the first half of 2025\, but often fail when presented with e ven medium time-horizon tasks. This talk presents a taxonomy of 4 traits of reasoning models -- skills\, calibration\, strategy\, and abstraction -- that will be crucial to creating the next generation of AI application s. With this\, we focus on the latter two\, strategy and abstraction\, a nd discuss how these traits will enable long-horizon and reliable agents. The talk concludes with a scenario where these agentic behaviors are the foundation for RL continuing to scale in the coming years and post-train ing techniques reaching compute parity with pretraining methors sooner th an later. DTEND:20250605T144000 DTSTAMP:20260403T174839Z DTSTART:20250605T142000 LOCATION:Yerba Buena Ballroom 2-6: Reasoning + RL SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:A taxonomy for next-generation reasoning models UID:SZSESSION926721 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Will Bryk\n\nRAG quality for AI agents is critical\, and traditional keyword-based search engines consistently underperform in agentic or multi-step tasks\, where semantic grounding and contextual nu ance matter most.\n\nIn this talk\, Will Bryk\, CEO of Exa will live code two AI agent applications–one using traditional keyword search RAG and o ne using neural network RAG via vector search. He’ll then evaluate both a pplications based on task performance\, relevance\, and latency. With a l ive demo (no theory or pre-baked applications)\, the audience will get a firsthand look at the practical differences between keyword and semantic systems in production\, and learn embedding strategies\, indexing trade-o ffs\, hybrid retrieval techniques\, prompt tuning\, and more. DTEND:20250605T144000 DTSTAMP:20260403T174839Z DTSTART:20250605T142000 LOCATION:Golden Gate Ballroom A: Retrieval + Search SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building a Smarter AI Agent with Neural RAG UID:SZSESSION914024 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: John Dickerson\n\nAI is getting deployed without guar drails\, without governance\, without due diligence. Surely this is the year we’ll see a Fortune 500 CEO fired because of a preventable AI incide nt. Surely this is the year we’ll see enterprises wake up to pre-deploym ent evaluation and post-deployment monitoring being an urgent need. This story hasn’t changed for a decade\, but surely this is the year it will. \n\nIn this talk\, I’ll cover what enterprise-level AI/ML evaluation has looked like for the last decade - what’s changed\, what hasn’t\, what sel ls\, what doesn’t\, and where I see things going from here on out. Evalu ation matters - we all know this - but using my experience in the trenche s over the last decade or so I hope to bridge the gap between what practi tioners need and what the C-suite pays for in the space of AI evaluations .\n DTEND:20250605T144000 DTSTAMP:20260403T174839Z DTSTART:20250605T142000 LOCATION:Golden Gate Ballroom B: Evals SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:2025 is the Year of Evals! Just like 2024\, and 2023\, and … UID:SZSESSION915059 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Jared Hanson\n\nWe all know sharing passwords is bad (unless you want free TV)\, so why are we sharing API keys with AI? We s houldn't\, and that’s why we need to talk about OAuth.\n\nIn this talk\, we will give a brief intro to OAuth. Then we will talk about the state o f authorization in MCP. We will show how an MCP client uses OAuth to aut henticate a user and securely access private resources and tools hosted b y an MCP server. Then we’ll look at ways autonomous agents can use OAuth on their own behalf\, talking to other agents and MCP servers directly. We’ll learn how to use OAuth to build agents that humans and machines ca n trust. DTEND:20250605T144000 DTSTAMP:20260403T174839Z DTSTART:20250605T142000 LOCATION:Foothill C: Security SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How to Secure Agents using OAuth UID:SZSESSION915751 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Steve Ruiz\n\nLearn about tldraw's latest experiments with AI on an infinite canvas. In 2024\, we created tldraw computer\, a loose visual programming environment where arrows and LLMs powered every step of a graph on tldraw's canvas. DTEND:20250605T144000 DTSTAMP:20260403T174839Z DTSTART:20250605T142000 LOCATION:Foothill G 1&2: Design Engineering SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:tldraw computer UID:SZSESSION934617 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Sharif Shameem\n\nCreating and sharing demos is the e asiest way to influence the future. It gets people to think about what's possible. A good tech demo doesn't have to be fully fleshed out. It doesn 't even have to be fully functional. The purpose of a demo is to inspire. A good demo makes you feel like someone jumped into the future and pulle d back an idea to the present. DTEND:20250605T144000 DTSTAMP:20260403T174839Z DTSTART:20250605T142000 LOCATION:Foothill F: Generative Media SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Good Demos are Important UID:SZSESSION947929 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Nikhil Abraham\n\nHow we converted a bimanual robot i nto a professional chef that works in novel kitchens and learn new recipe s from a single demonstration. DTEND:20250605T144000 DTSTAMP:20260403T174839Z DTSTART:20250605T142000 LOCATION:Foothill E: Autonomy + Robotics SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:General purpose robots as professional Chefs UID:SZSESSION945538 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Rita Kozlov\n\nAI workloads are rapidly shifting from AI being used for augmentation (co-pilots)\, to AI becoming responsible for full\, end-to-end automation (agents). But building effective agents\ , and even more importantly\, agent experiences that boost productivity r equires many pieces. In this talk\, we'll be covering the building blocks of agents\, how to put them together\, and what we've learned from top c ompanies building agents along the way. DTEND:20250605T150000 DTSTAMP:20260403T174839Z DTSTART:20250605T144000 LOCATION:Golden Gate Ballroom C: AI in the Fortune 500 SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Agents (the hard parts!) UID:SZSESSION935969 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Paul Klein IV\n\nWith the rise of MCP servers\, A2A\, and our trusty friend\, OpenAPI\, it turns out the web browser may be th e default MCP server for the rest of the internet.\n\nIn this talk\, we'l l walk through how a web browsing tool is probably the only tool you'll n eed to enable production AI Agents. DTEND:20250605T150000 DTSTAMP:20260403T174839Z DTSTART:20250605T144000 LOCATION:SOMA: AI Architects SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The Web Browser Is All You Need UID:SZSESSION914934 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Josh Albrecht\n\nIn this case study-based keynote\, J osh Albrecht\, CTO of Imbue\, examines the critical engineering challenge s in building AI coding systems that create more than just prototypes. Dr awing from Imbue's research developing Sculptor\, an experimental coding agent environment\, Josh shares key insights into the fundamental technic al obstacles encountered when evolving AI-assisted coding from toy applic ations to more robust software systems. \n\nThe session will explore appr oaches to core challenges like safely executing code\, managing context a cross large codebases\, automating test generation\, and creating systems that can identify potential pitfalls in AI-generated code. Attendees wil l gain practical insights into the technical underpinnings of next-genera tion coding agents that aim to handle complex software engineering challe nges architecting larger systems\, increasing meaningful test coverage an d designing systems that are easy to debug—moving us closer to AI systems that can help create maintainable software.\n DTEND:20250605T150000 DTSTAMP:20260403T174839Z DTSTART:20250605T144000 LOCATION:Yerba Buena Ballroom 7&8: SWE Agents SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Beyond the Prototype: Using AI to Write High-Quality Code UID:SZSESSION914017 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Christian Szegedy\n\nI describe a new paradigm toward s open-endedly self-improving intelligence by scaling verification to rem ove the human data and supervision bottleneck. The objective is to achiev e trustless alignment of superintelligence. DTEND:20250605T150000 DTSTAMP:20260403T174839Z DTSTART:20250605T144000 LOCATION:Yerba Buena Ballroom 2-6: Reasoning + RL SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Towards Verified Superintelligence UID:SZSESSION939640 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: David Karam\n\nStart with the simplest Search - in-me mory embeddings with relevance ranking. End with the most complex planet- scale Search - 70+ corpus mix of token\, embeddings\, and knowledge graph s\, all jointly retrieved\, custom ranked\, joint re-ranked\, and then LL M-processed\, at 160\,000 queries per second in under 200msec.\n\nThis ta lk will be a fun “one query at a time” survey of all techniques in RAG in incremental complexity\, showing the limits of each technique and what t he next layered one opens up in terms of capabilities to handle ever-more complex queries in RAG. You’ll learn why queries like [falafel] are noto riously hard to Search over\, why chunking your documents can be disastro us\, how you can sometimes can get away with a simple bm25\, and how some Search problems are so hard to solve that you’re better off punting the problem to the LLM or the UX. Brought to you by the team that worked on 5 0+ Search products\, in the context of Google.com and custom Enterprise S earch. DTEND:20250605T150000 DTSTAMP:20260403T174839Z DTSTART:20250605T144000 LOCATION:Golden Gate Ballroom A: Retrieval + Search SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Layering every technique in RAG\, one query at a time UID:SZSESSION907695 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Jason Liu\, Jeff Huber\n\nBy the end of this talk\, you'll understand what it takes to apply clustering techniques and data a nalysis to understand what is the valuable work that your AI application is doing through analyzing conversation histories and how to create gener ative evals to benchmark your newly discovered superpowers. DTEND:20250605T150000 DTSTAMP:20260403T174839Z DTSTART:20250605T144000 LOCATION:Golden Gate Ballroom B: Evals SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How to look at your data\; what to look for\, how to measure UID:SZSESSION916104 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Rene Brandel\n\nWe hacked 7 of the16 publicly-accessi ble YC X25 AI agents. This allowed us to leak user data\, execute code re motely\, and take over databases. All within 30 minutes each. In this ses sion\, we'll walk through the common mistakes these companies made and ho w you can mitigate these security concerns before your agents put your bu siness at risk. DTEND:20250605T150000 DTSTAMP:20260403T174839Z DTSTART:20250605T144000 LOCATION:Foothill C: Security SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:How we hacked YC Spring 2025 batch’s AI agents UID:SZSESSION915873 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Craig Wattrus\n\nDesigning user experiences for AI me ans moving beyond traditional interfaces.\n\nDesigners are grappling with how to create intuitive and effective interactions for these new AI capa bilities\, while growing their practice to include philosophy\, ethics an d coding. \n\nWhat if AI interactions could be reimagined as new 'coworke rs'? This talk explores AI systems as your new coworkers. Covering novel UX patterns we’ve implemented and are researching at Flatfile as well as a state of the union on emergent patterns we’re seeing and using from the industry.\n\nAttendees will get a peek into explorations into AI cursors \, forward-leaning chat paradigms and tool UX. We will discuss both work thats in production today at some of our biggest customers as well as tho ught-provoking demos\, offering a vision for the future of AI UX. DTEND:20250605T150000 DTSTAMP:20260403T174839Z DTSTART:20250605T144000 LOCATION:Foothill G 1&2: Design Engineering SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Form factors for your new AI coworkers UID:SZSESSION915783 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Mark Bissell\n\nThe goal of mechanistic interpretabil ity is to reverse engineer neural networks. Having direct\, programmable access to the internal neurons of models unlocks new ways for developers and users to interact with AI — from more precise steering to guardrails to novel user interfaces. While interpretability has long been an interes ting research topic\, it is now finding real-world use cases\, making it an important tool for AI engineers. DTEND:20250605T150000 DTSTAMP:20260403T174839Z DTSTART:20250605T144000 LOCATION:Foothill F: Generative Media SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Why you should care about AI interpretability UID:SZSESSION914798 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: JingXiang Mo\n\nIntroducing developer ready robots th at are open-source\, affordable\, and easy to use. DTEND:20250605T150000 DTSTAMP:20260403T174839Z DTSTART:20250605T144000 LOCATION:Foothill E: Autonomy + Robotics SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Scaling Open-source Humanoid Robots UID:SZSESSION948652 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250605T154500 DTSTAMP:20260403T174839Z DTSTART:20250605T150000 LOCATION:Atrium: Event Hub SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Afternoon Break UID:SZSESSION5d3d0ffe-26d7-4142-9319-ea97c2f2acd4 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ethan Sutin\n\nYour smartphone knows your location\, your smartwatch tracks your heartbeat\, but what if AI could understand y our entire life context? The idea is obvious (just ask Sam and Jony)\, bu t the reality of ambient intelligence brings technical challenges no one talks about\, from processing human context at scale to making it actuall y useful.\n\nBut what if we could crack the code on truly personal AI tha t lives with you\, not just on your phone? Enter the era of ambient perso nal intelligence.\n\nWe'll dive into hard-won lessons from processing ove r 150 billion tokens of personal context. We will discuss privacy-first s ystems from edge computing to Secure Enclaves\, discover why ambient unde rstanding is both harder and more powerful than you think\, and explore t he frontier where personal AI agents continuously reason about your needs and take actions proactively DTEND:20250605T151500 DTSTAMP:20260403T174839Z DTSTART:20250605T150000 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The Buzz About Ambient Personal AI: What Really Works UID:SZSESSION948075 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Henry Weller\n\nThis talk will explore typical and fo rward-looking use cases for Atlas Vector Search\, as well as how differen t types of data models and query patterns can be implemented and effectiv ely scaled to meet the needs of those use cases. There will be a focus on the "Iron Triangle of Search" balancing accuracy\, speed\, and cost and talking about practical considerations that emerge within those use cases .\n\n\nThis will be a technical talk focused on the "how" of Atlas Vector Search and considerations when building information retrieval systems gi ven by a technical PM\, not a sales pitch explaining how basic vector ret rieval "solves" hallucinations. DTEND:20250605T151500 DTSTAMP:20260403T174839Z DTSTART:20250605T150000 LOCATION:Salons 9-15: Expo Hall SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Building Vector Search Experiences with MongoDB: Access patterns\, data models\, and scaling considera UID:SZSESSION914371 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Eno Reyes\n\nPlanning\, coding\, testing\, monitoring —the endless cycle that spans 10+ tools that fragment our focus and slows delivery to a crawl. Vibe coding doesn't work when you've got 10TB of co de. If you just sighed\, you're one of many professional software enginee rs trapped in the traditional software development lifecycle (SDLC) that was designed before AI could parallelize your entire workflow.\n\nBut wha t if you could orchestrate multiple AI agents on tasks beyond just genera ting code\, while you focus on the creative decisions that matter?\n\nIn this talk\, I'll demonstrate how real enterprise organizations are changi ng their entire SDLC—going from understanding\, planning\, coding\, and t esting all the way to incident response—using AI agents. You'll witness t he next evolution of software engineering—where AI doesn't just generate code\, but orchestrates the entire development lifecycle. DTEND:20250605T152000 DTSTAMP:20260403T174839Z DTSTART:20250605T150000 LOCATION:Yerba Buena Ballroom 7&8: SWE Agents SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Ship Production Software in Minutes\, Not Months UID:SZSESSION904751 END:VEVENT BEGIN:VEVENT DESCRIPTION:[last round of Attendee-Led 10min lightning talks]\n\nCursor f or Staff Engineers: Learn from someone who's completely transitioned to A I-assisted coding\, addressing common concerns like "I could write it fas ter manually\," "I know my codebase too well to need help\," and integrat ion challenges—this session delivers practical workflows that work with r eal systems.\nVignesh Mohankumar DTEND:20250605T152500 DTSTAMP:20260403T174839Z DTSTART:20250605T151500 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Cursor for Staff Engineers — Vignesh Mohankumar UID:SZSESSIONe642e026-10a8-4410-95df-e093676d76a3 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Forrest Brazeal\, Matt Ball\n\nShortened presentatio n-only version of our Apollo 11 workshop DTEND:20250605T153000 DTSTAMP:20260403T174839Z DTSTART:20250605T151500 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:To the moon! Navigating deep context in legacy code with Augment A gent UID:SZSESSION936298 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Kenneth DuMez\n\nAs magical as they may seem\, AI age nts should be treated like any other software system. This talk will cove r the best practices in designing and building AI systems including obser vability\, security hardening\, and proper UX. DTEND:20250605T153000 DTSTAMP:20260403T174839Z DTSTART:20250605T151500 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Cattle\, not genies: building AI agents from first principles UID:SZSESSION933474 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Nathan Sobo\n\nSoftware engineers have long understoo d that high-quality code requires comprehensive automated testing. For de cades\, our industry has relied on deterministic tests with clear pass/fa il outcomes to ensure reliability. \n\nHigh-quality software depends on a utomated testing. That's certainly true at Zed\, where we're building a n ext-generation native IDE in Rust. Zed runs at 120 frames per second\, bu t it would also crash once a second if we didn't maintain and run a compr ehensive suite of unit tests on every change.\n\nBut what happens when AI enters the equation?\n\nIn this talk\, we'll explore how continuous inte gration evolves when working with AI components. "Evals" - parlance from the machine learning field - are fundamentally a continuation of the soft ware testing tradition\, but with a critical difference: they're inherent ly stochastic.\n\nZed's traditional CI goes to extreme lengths to elimina te non-determinism\, as nobody likes having their pull requests blocked b y flaky builds. We've even fully simulated network interactions with a de terministic random scheduler. AI components\, however\, forced us to conf ront a fundamental paradigm shift—uncertainty isn't a bug but an intrinsi c feature of these systems\, compelling us to embrace what we couldn't av oid.\n\nWe'll share our journey of reconceptualizing evals as "stochastic unit tests" - still verifying system behavior\, but without binary pass/ fail grades.\n\nWe'll discuss practical approaches to:\n- Thoughtfully bu ilding test suites for AI components\n- Shifting from red/green outcomes to "shades of gray"\n- Replacing build gates with trend analysis and perf ormance monitoring\n- Maintaining engineering confidence despite statisti cal variance\n\nWhether you're incorporating AI into existing systems or building new AI-powered tools\, this talk will provide practical insights into maintaining quality when determinism gives way to probability. DTEND:20250605T153000 DTSTAMP:20260403T174839Z DTSTART:20250605T151500 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:CI in the Era of AI: From Unit Tests to Stochastic Evals UID:SZSESSION915826 END:VEVENT BEGIN:VEVENT DESCRIPTION:[last round of Attendee-Led 10min lightning talks]\n\nNever Ha llucinate - A lightweight documentation framework for ensuring coding age nts never hallucinate - even across multiple coding sessions DTEND:20250605T154000 DTSTAMP:20260403T174839Z DTSTART:20250605T153000 LOCATION:Grand Assembly SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Never Hallucinate — Barada Sahu UID:SZSESSION83bb55a7-899a-47db-80b2-96c73c774d24 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Sam Fertig\n\nSometimes it seems like Windsurf knows you a little too well. It's one thing to generate generic code\, but to p redict your next intent? From matching existing code patterns and styles to tracking how local changes affect the larger codebase\, this talk digs into the technical challenges of context awareness and why simply indexi ng code falls short. Relive our journey tackling the core issue in the AI IDE space : balancing retrieval quality with latency constraints and sca ling effectively as codebases grow. For those curious about the infrastru cture behind context-aware AI\, this talk offers insights into our approa ch of turning massive codebases into collections of useful context. DTEND:20250605T154500 DTSTAMP:20260403T174839Z DTSTART:20250605T153000 LOCATION:Nobhill A&B: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The Eyes Are The (Context) Window to The Soul: How Windsurf Gets t o Know You UID:SZSESSION933633 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Beyang Liu\n\nIt's raining coding agents. But while m any are saying they're feeling the AGI\, others say they're not that usef ul for serious programming. How much is hype and how much is a skill issu e? We'll share empirical observations that help explain the divergence of developer opinion. And we'll cover emergent strategies uncovered by user s of Amp\, a new coding agent in research preview\, that can help you emp loy agents to complete more complex tasks in production codebases. DTEND:20250605T154500 DTSTAMP:20260403T174839Z DTSTART:20250605T153000 LOCATION:Willow: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The emerging skillset of wielding coding agents UID:SZSESSION933575 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Emma Ning DTEND:20250605T154500 DTSTAMP:20260403T174839Z DTSTART:20250605T153000 LOCATION:Juniper: Expo Sessions SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Empowering Developers to build Cutting-Edge AI experiences on devi ce UID:SZSESSION936907 END:VEVENT BEGIN:VEVENT DESCRIPTION: DTEND:20250605T160000 DTSTAMP:20260403T174839Z DTSTART:20250605T154000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Keynote Doors UID:SZSESSIONd9911bf6-c869-469a-8e47-8258f6bbcfbf END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: George Cameron\, Micah Hill-Smith\n\nThe entire AI s tack is developing faster than ever - from chips to infrastructure to mod els. How do you sort the signal from the noise? Artificial Analysis an in dependent benchmarking and insights company dedicated to helping develope rs and companies pick the right models and technologies for building appl ications. This talk will walk through the state of the frontier across th e AI stack. DTEND:20250605T162000 DTSTAMP:20260403T174839Z DTSTART:20250605T160000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Trends Across the AI Frontier UID:SZSESSION936564 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Ankur Goyal\n\nThe final word on Evals DTEND:20250605T162500 DTSTAMP:20260403T174839Z DTSTART:20250605T162000 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Evals Closing Keynote UID:SZSESSION943899 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Barr Yaron\n\nCome hear the results of the 2025 State of AI Engineering. DTEND:20250605T163500 DTSTAMP:20260403T174839Z DTSTART:20250605T162500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:State of AI Engineering 2025 UID:SZSESSION943904 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Alex Atallah\n\nHow the first LLM aggregator got star ted\, some of the weird moments in its early growth\, architecture challe nges\, and where we'll be taking it down the road DTEND:20250605T165500 DTSTAMP:20260403T174839Z DTSTART:20250605T163500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:fun stories from building OpenRouter and where all this is going UID:SZSESSION941906 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speaker: Sean Grove\n\nIn an era where AI transforms software development\, the most valuable skill isn't writing code - it's communica ting intent with precision. This talk reveals how specifications\, not pr ompts or code\, are becoming the fundamental unit of programming\, and wh y spec-writing is the new superpower.\n\nDrawing from production experien ce\, we demonstrate how rigorous\, versioned specifications serve as the source of truth that compiles to documentation\, evaluations\, model beha viors\, and maybe even code. \n\nJust as the US Constitution acts as a ve rsioned spec with judicial review as its grader\, AI systems need executa ble specifications that align both human teams and machine intelligence. We'll look at OpenAI's Model Spec as a real-world example.\n\nFinally\, w e'll end on some open questions about what the future of developer toolin g looks like in a world where communication once again becomes the most i mportant artifact in engineering. DTEND:20250605T171500 DTSTAMP:20260403T174839Z DTSTART:20250605T165500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:The New Code UID:SZSESSION925974 END:VEVENT BEGIN:VEVENT DESCRIPTION:Speakers: Benjamin Dunphy\, Shawn Wang DTEND:20250605T173000 DTSTAMP:20260403T174839Z DTSTART:20250605T171500 LOCATION:Keynote/General Session (Yerba Buena 7&8) SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:AI Engineer World's Fair Hackathon - Grand Prize UID:SZSESSION936046 END:VEVENT BEGIN:VEVENT DESCRIPTION:Organize your own in the app/AIE slack\, or sign up at https:/ /www.ai.engineer/#events DTEND:20250605T210000 DTSTAMP:20260403T174839Z DTSTART:20250605T173000 LOCATION:Atrium: Event Hub SEQUENCE:329002 STATUS:CONFIRMED SUMMARY:Community Meetups (Jun 5) UID:SZSESSION9841fa43-2756-4c5a-8f87-b745d81bad92 END:VEVENT END:VCALENDAR