Kiruthika Subramani
Google Developer Expert
Montréal, Canada
Actions
Kiruthika Subramani is an aspiring AI professional with a robust foundation in developing and implementing AI solutions. She works as a Data Scientist in Bell Canada. She holds industry certifications in Data, Machine Learning, and Cloud, and is recognized as an IBM Dual Champion and Google Developer Expert. Kiruthika is a prolific speaker and author, having delivered over 200 AI talks and published 100 blogs. She is committed to driving innovation and ethical AI practices, ensuring responsible and impactful advancements in the field.
Kiruthika is currently pursuing a Master’s in Computer Science at Université de Montréal (MILA), Canada, and holds a B.Tech in Artificial Intelligence and Data Science from M.Kumarasamy College of Engineering, Karur, India. Her technical skills span across programming languages (Python, Java, C++), machine learning, deep learning, NLP, computer vision, responsible AI, cloud platforms, and various tools and frameworks.
Authored two books on AI and holds several professional certifications. She has been recognized as a Google Developer Expert, IBM Dual Champion, and Google Cloud Champion Innovator. Her key accomplishments include winning multiple hackathons and receiving the Best Outgoing Student Award. She has also been spotlighted as Today’s Architects by IBM and has delivered numerous global tech talks and authored a popular blog series on Medium
Area of Expertise
Topics
The Art of Storytelling with Data
The Art of Storytelling with Data is designed to help participants choose the right visualizations to convey their message effectively. The session covers various visualization types, such as bar charts, line graphs, and scatter plots, explaining when and why to use each one to tell a compelling story. It also delves into the principles of visual storytelling, including how to highlight key insights, avoid common pitfalls, and create emotionally engaging narratives. Participants will learn to tailor their visualizations to different audiences, ensuring clarity and impact. This non-technical session is ideal for anyone looking to transform data into captivating stories that resonate with a broad audience.
Connect, communicate, and extend agents with MCP, A2A, and Gemini Skills
We all know how to build MCP servers by now. If not, no worries, we cover that too. But did you know your tools need proper authentication? Most MCP servers out there have no auth layer at all and that is a security gap sitting in production. In this session you learn how MCP Auth locks down your tool connections and how Remote MCP moves your servers to the cloud so they are not stuck running on someone's laptop.
Then we go further. What happens when your agent needs help from another agent on a completely different team? That is A2A. Agents discover each other through Agent Cards, negotiate what they can do, and delegate tasks across system boundaries without anyone writing custom glue code.
Then there is Gemini Skills. Instead of building every capability from scratch, you give your agent pre-built reusable skills that let it handle specific tasks out of the box, think of them like plugins your agent can use to instantly get better at things without you writing the logic yourself. You leave knowing how all three protocols connect, when to use which one, and with working code that wires them together.
Build AI agents your compliance team will actually approve
Your agent is leaking PII, recommending competitors, and costing you triple.
Your agent is live. A customer asks a product question and your agent says "actually, competitor has a better deal." Your logs are full of customer phone numbers and emails in plain text, a GDPR and PIPEDA violation nobody on the team noticed. The same question asked twice gives two different answers because the prompt has no structure. Your monthly bill is 3x what you estimated because nothing limits how many tokens a single conversation burns. These are not edge cases. This is what most agents in production look like right now.
Attendees build a production-ready agent from scratch using LangChain and Google ADK. A middleware layer that scans every message for PII and redacts it before anything reaches the model or the logs. Guardrails that block competitor mentions, off-topic answers, and hallucinated claims. Structured prompts so the same input produces consistent output regardless of phrasing. Token budgets and response caching so cost stays predictable. Async tool calls so the agent stays responsive on slow APIs. The finished agent deploys to Google Cloud Run with Secret Manager handling credentials. Attendees leave with the full codebase and a pattern they can reuse on every agent after this.
One user works, a thousand don't - Scaling AI products on Google Cloud
Your AI product works perfectly, for one user at a time. In this hands-on demo, we break a working agent with concurrent traffic, then rebuild it to scale. We deploy to Cloud Run for automatic horizontal scaling, move session state to Firestore so any instance can serve any user, add Gemini context caching and token budgets to keep costs flat as traffic grows, and wire up Cloud Monitoring to watch it all live. Then we load-test it on stage and watch it hold. Bring your sprint project , this is the architecture it needs before launch day.
From demo to production - Auto-scale your AI agent on Cloud Run
You shipped an AI product in four weeks. Now what happens when 100 users hit it at once? In this live demo workshop, we take an agent built with Google ADK and deploy it to Google Cloud the right way. We containerize it, push it to Cloud Run with auto-scaling and concurrency tuning, add Firestore for session memory so state survives across instances, and load-test it live to watch instances spin up under real traffic. You leave knowing the exact architecture to take your sprint project from laptop demo to production without rewriting it.
Model Chaining for Complex AI Systems
In this session, you will understand how to build complex AI systems using model chaining techniques. I’ll guide you through linking multiple AI models to perform sophisticated tasks, with lab exercises designed to help you create and deploy your own model chains effectively.
Let us build RAG that sees charts, connects documents, and stops ignoring half your data
Your RAG reads text. That is it. Someone uploads a PDF with a table, a chart, a diagram and your RAG skips all of it. The answer was right there in the image and your agent acted like it did not exist. Then someone asks a question that needs information scattered across three different documents. A name in one, a date in another, a policy in a third. Your RAG cannot connect them because it treats every chunk like it lives alone with no relationship to anything else.
In this session you learn three types of RAG and build each one. Text RAG done right with semantic chunking, reranking, and metadata filtering so you retrieve the right chunk not just a similar one. Multimodal RAG using Gemini so your agent actually sees the charts, images, tables, and diagrams it was ignoring before. Graph RAG where entities and relationships across documents are stored in a knowledge graph so your agent connects facts across hundreds of files instead of searching one chunk at a time. You build all three on real messy data, see where each one wins, where each one breaks, and learn when to use one vs combine all three.
Let us build AI agents with real memory, real personalization, and real feedback loops
Your agent treats every conversation like a first date. It should not.
You ask your agent the same thing every Monday. It has no idea. You correct it and it makes the same mistake next time. You told it your preferences three conversations ago and it forgot everything. Most agents have no memory beyond the current chat window. Every session starts from zero.
In this session you learn how to build agents that actually learn from the people using them. You start with short-term memory so the agent stays coherent within a conversation. Then long-term memory so it remembers what happened days or weeks ago using vector stores over past conversations. Then user preference tracking so the agent learns how you like things done and adapts its tone, format, and decisions over time. Then feedback loops where the agent takes your corrections and actually uses them to improve future responses instead of ignoring them. You see how to store all of this safely, how to let users see and delete what the agent remembers, and how to handle the privacy side so personalization does not become surveillance. You leave with an agent that gets better the more you use it and respects your data while doing it.
Let us run AI models on your laptop and your browser with zero cloud cost
Session abstract
What if your AI never needed an API key, a cloud account, or an internet connection?
Every AI app you have used sends your data to someone else's server. Every token costs money. Every request needs internet. For a lot of use cases that is completely unnecessary. Gemma runs on your laptop. It runs in your browser. Your data never leaves your machine and your bill is zero.
In this session you learn three ways to run AI without the cloud. First, Ollama on your laptop with Gemma. You pull the model, run it locally, and build an app on top of it that works offline. Second, Chrome built-in AI with Gemini Nano running directly in the browser. No server, no API, the model lives in the browser tab and you build a web app that uses it. You see what tasks each setup handles well, where the quality drops compared to a full cloud model, and when local is genuinely the better choice. You leave knowing when to use which setup, what the quality tradeoffs are, and with working code for both that you can run on your own machine today.
Build the same AI agent in six frameworks and learn which one actually fits your problem
Everyone has a favourite framework. Nobody has tried all of them on the same problem.
You learned LangChain from a tutorial. Your colleague swears by ADK. Someone says AgentKit is the future. CrewAI keeps showing up in multi-agent demos. And your manager asks why you are writing code when Vertex AI Agent Builder exists. Each one solves a different problem and picking the wrong one means weeks fighting the framework instead of building your product.
Attendees build the same agent across LangChain, ADK, AgentKit, LangGraph, and CrewAI on the same task. They add tools, memory, and error handling in each and feel the difference. They see where Vertex AI Agent Builder solves it in minutes and where it hits limits. Every framework gets wired into LangSmith so attendees can trace agent decisions, compare latency, inspect tool calls, and evaluate output quality across all six in one dashboard. The comparison is not vibes, it is data. You leave knowing which framework fits which problem, what each one costs you, and what LangSmith reveals about them that the docs never will.
Build multi-agent AI systems with the right pattern for the right problem
Your team has decided they need a multi-agent system. That was probably the right call. But now everyone is spending time figuring out whether we need three agents or five agents. That is the wrong question. The number of agents is an outcome, not a decision. It comes from the pattern you choose, not from counting tasks on a whiteboard.
In this session you learn to pick the pattern first and let the agents follow. A supervisor that routes tasks to specialists. A sequential pipeline where agents hand off in order. A hierarchical tree where agents manage their own sub-agents. Role-based teams where each agent has a clear job. Controlled loops where an agent refines its own output but knows when to stop. Each pattern tells you exactly how many agents you need and what each one does. You learn how they share state without overwriting each other and how to keep one failure from crashing the whole system. Then you hit the point every growing team hits. Your agents are no longer in one codebase. The research agent lives with the data team. The writing agent lives with the content team. Different systems, different people. That is where A2A comes in. Agents discover each other, negotiate capabilities, and delegate tasks across team and system boundaries. You also learn how to trace every decision across every agent so when something breaks you see exactly where and why.
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top