Welcome - 9:00 AM – 9:10 AM PST

Session 1: AI Sprint: Building interactive agentic applications using ADK and AG-UI protocol
Learn how to build interactive, real-world agentic applications by combining Google’s Agent Development Kit (ADK) with the AG-UI protocol. I focus on the practical challenges that arise when moving from simple chatbots to multi-step, tool-calling AI agents, such as managing state, handling asynchronous workflows, and creating user interfaces that can clearly reflect what agents are doing in real time.

Break - 10:15 AM – 10:25 AM PST

Session 2: LLM Interpretability & Open Source Frameworks
This session explores the foundational concepts of building responsible, production-ready agent tools using Google’s open-source ecosystem.

This session dives deep into the "black box" of Large Language Models to explore how we can understand and steer model behavior at a foundational level. We will focus on Gemma Scope, a comprehensive library of Sparse Autoencoders (SAEs) trained on the internal layers of the Gemma 2 model family.

Attendees will learn to use these SAEs to decompose complex neural activations into human-interpretable "features," effectively mapping the model's thought process. By moving beyond simple prompting and into the realm of mechanistic interpretability, we will demonstrate how to identify specific neurons responsible for concepts, biases, and safety triggers. This showcases the tools to audit model internals, ensuring that agentic systems perform well for the right reasons.

Wrap up & Network — 11:25 AM – 12:00 PM PST (Optional)