The Evolution of Video RAG

Video RAG used to mean frame extraction, audio separation, multiple embeddings, and complex orchestration pipelines.

A year ago, if you wanted to build semantic search over video content, you'd extract frames, transcribe audio separately, generate embeddings for each modality, calculate similarities, and orchestrate everything yourself. Three hundred lines of Python. Six different libraries. Multiple failure points.

Then multimodal models that natively understand video arrived—and everything changed.

I'll show you both approaches through working code:

The Traditional Pipeline:
- Frame extraction and key frame selection using cosine similarity
- Audio transcription with speaker diarization and timestamps
- Separate embedding generation for visual and audio content
- Vector storage and semantic search with `pgvector`
- Custom orchestration logic to tie it all together

The Modern Approach :
- Native video understanding without decomposition
- Unified temporal embeddings across modalities
- Agent-based architectures that handle orchestration
- Production-ready patterns in a fraction of the code

Join me for a live demonstration of native video understanding without the need to create embedding pipelines. Build a complete video analysis agent in minutes.

Graph-RAG vs RAG: Stop Inaccurate AI Agent Responses

Traditional RAG has a fundamental limitation: vector search retrieves text, not structured data. When you need precise answers, LLMs guess aggregations from text chunks
instead of executing calculations. This causes wrong averages, fabricated counts, and approximate results.

Graph-RAG solves this by storing data as entities and relationships, enabling structured queries that execute calculations instead of estimating from text. Join me as I build
a travel agent demo that compares both on 515K hotel reviews proving where RAG fails and how graphs deliver accurate answers.

Build multi-modal AI agents with scalable chat memory

The open-source Strands Agents SDK enables building production-ready, multi-agent AI systems in a few lines of code. The session demonstrates creating custom tools for video content analysis and converting them into Model Context Protocol (MCP) servers. Live code examples show how to implement scalable chat memory with Amazon S3 Vectors, from local development to enterprise production. Developers will learn how agents maintain context across conversations and provide personalized responses through efficient memory storage and retrieval.

Developers will learn how multimodal agents maintain context across conversations and provide personalized responses through efficient memory storage and retrieval

Building intelligent research agents for production environments

Modern applications increasingly need to access and synthesize real-time information from the web, but building intelligent research capabilities from scratch presents significant challenges. Developers struggle with API management, credential security, conversation context, and integrating multiple AI services into cohesive workflows.

This talk demonstrates how to build production-ready research agents in minutes, that provides secure runtime, identity management, and API gateway capabilities for AI agents. You'll learn to create agents that can conduct iterative web research, maintain conversation context, and provide structured responses with proper source attribution.

This talk demonstrates how to build production-ready research agents. You'll learn to create agents that can conduct iterative web research, maintain conversation context, and provide structured responses with proper source attribution.

From Prototype to Production in Minutes: Deploying AI Agents That Scale

The gap between building an AI agent prototype and deploying it to production costs companies months of engineering time and thousands of dollars. This talk bridges that gap, showing how to deploy production-grade AI agents in minutes instead of months.

Using modern agentic frameworks and cloud infrastructure, you can deploy enterprise-ready agents with:

Cross-session memory for personalized experiences
Zero-code monitoring for observability
Auto-scaling infrastructure that grows with demand
Cost optimization patterns from real production workloads

Reduce AI Agents Costs and Mistakes with Semantic tool Selection

AI agents with many tools face a dual problem: they pick the wrong tool and waste tokens because tool descriptions get serialized into the context on every call. As agents scale to 50+ tools, errors increase and costs explode.

Semantic tool selection filters tools before they reach the LLM context using vector search, reducing errors and token costs.

This talk walks you through building TWO versions of the same travel agent to prove semantic tool selection works. By comparing the agents side-by-side on identical queries, you'll see dramatic improvements in both accuracy and cost and learn how to implement dynamic tool filtering that preserves conversation memory across multi-turn sessions.

You'll leave with working Python code and production patterns you can deploy immediately.

AI agents waste tokens sending all tool descriptions on every call and pick wrong tools as they scale. Learn how semantic tool selection reduces errors 75% and token costs 89% using vector search. Live demo with production-ready code you can use today.

Graph-RAG vs RAG: Stop Inaccurate AI Agent Responses

Traditional RAG has a fundamental limitation: vector search retrieves text, not structured data. When you need precise answers, LLMs guess aggregations from text chunks
instead of executing calculations. This causes wrong averages, fabricated counts, and approximate results.

Graph-RAG solves this by storing data as entities and relationships, enabling structured queries that execute calculations instead of estimating from text. Join me as I build
a travel agent demo that compares both on 515K hotel reviews proving where RAG fails and how graphs deliver accurate answers.

Traditional RAG has a fundamental limitation: vector search retrieves text, not structured data. When you need precise answers, LLMs guess aggregations from text chunks
instead of executing calculations. This causes wrong averages, fabricated counts, and approximate results.

Stop AI Agents from Making Up Facts: Graph-RAG for Reliable Data Retrieval

AI agents don't just answer wrong—they act wrong. When agents hallucinate during execution, they
fabricate API parameters, invent confirmations, and make decisions based on false data. Traditional
RAG retrieval makes this worse by returning approximate matches instead of precise answers. In
this hands-on session, we'll build a travel booking agent and compare two approaches: traditional
vector search (FAISS) vs Graph-RAG (Neo4j knowledge graph) on 300 hotel FAQ documents. You'll see
how structured knowledge graphs eliminate fabricated statistics and incomplete retrieval—achieving
better accuracy on precise queries. Walk away with a working pattern you can apply to any domain
where AI agents need reliable, verifiable data retrieval. No prior graph database experience
needed.

Smarter AI Agents: Semantic Tool Selection to Cut Errors and Reduce Costs

When AI agents have dozens of similar tools, they pick the wrong one—and you pay for it. Research
shows tool selection hallucinations increase with tool count, causing agents to call non-existent
functions, choose semantically wrong tools, or bypass tools entirely. In this session, we'll build
a travel agent with 31 tools and use vector-based semantic filtering (FAISS + SentenceTransformers)
to dynamically select only the most relevant tools per query. You'll see how filtering
dramatically improves accuracy while significantly reducing token costs per call. This technique
works with any AI agent framework and scales to hundreds of tools. You'll leave with a reusable
pattern for intelligent tool filtering that makes your agents faster, cheaper, and more reliable.

Beyond Prompt Engineering: Neurosymbolic Guardrails AI Agents Cannot Bypass

Prompt engineering cannot enforce business rules. Agents ignore docstring instructions, confirm bookings without payment, and accept invalid parameters—because prompts are suggestions, not constraints.

In this session, I'll compare two approaches: prompt-only rules vs neurosymbolic integration where symbolic validation executes at the tool level.

I'll build a travel booking agent where business rules (payment verification, guest limits, prerequisite checks) are enforced through verifiable code the LLM physically cannot bypass.
The result: consistent business rule compliance regardless of how creative the prompt.

Walk away with a practical pattern for combining LLM
flexibility with deterministic rule enforcement, applicable to any domain with compliance requirements.

You'll leave with working Python code and production patterns you can deploy immediately.

Catching AI Agentic Hallucinations with Multi-Agent Validation

Single AI agents hallucinate without detection claiming success when operations fail, using wrong
tools, and fabricating responses. No self correction mechanism exists in isolation. In this
session, we'll build a multi agent validation system using the Executor → Validator → Critic
pattern where specialized agents cross validate each other's work through structured debate. Using
Strands Agents Swarm orchestration, you'll see how this pattern catches invalid operations, wronga
tool usage, and fabricated responses before they reach users. Based on research showing multi-agent
debate significantly reduces hallucination rates, you'll leave with a working cross validation
architecture applicable to any high stakes AI agent deployment.

Context Engineering - Stop Agents from Choking on Their Own Data

AI agents choke when tools return large outputs logs, database results, file contents. The data fills the context window, older messages get pushed out, and the agent silently produces incomplete results. Worse: every subsequent operation reprocesses the same data, wasting tokens.

Memory pointers solve both problems: tools store large data outside context and return short references. The LLM orchestrates analysis without ever seeing the raw data.

Join me as I demo this pattern in single-agent and multi-agent configurations with production-ready code.

You'll leave with working Python code and production patterns you can deploy immediately.