| 🌟 Speaker Connie Leung: Private and Offline: On-Device AI with LiteRT.js and Gemma 4 | Get hands-on with LiteRT.js and Google’s open-source Gemma 4 models. Learn to retrieve models from Hugging Face and leverage WebGPU for high-performance in-browser inference.
Key Takeaways: Setting up LiteRT-LM; optimizing WebGPU for browser acceleration; designing offline-first security. |
| 🛠️ Speaker Jay Thakkar : Gemma in Production: Enterprise-Grade, Secure & Scalable LLM Deployment on Cloud Run | Discover how to take Google’s open-source Gemma LLM from experimentation to enterprise production with confidence. This session explores secure deployment strategies using Cloud Run, enabling scalable, serverless inference without compromising sensitive data.
Learn how Gemma empowers organizations to retain full control over model behavior and data privacy—eliminating risks of external training or data leakage—while delivering high-performance AI at scale. Ideal for teams building compliant, trust-centric GenAI applications in regulated environments. |
| 🛠️ Speaker Lesly Zerna: From Chatbot to Mentor: Building a Local Socratic Tutor with Gemma 4 | Most AI applications suffer from "answer-dumping"—they do the homework for the user instead of teaching them. In this practical, hands-on session, we will build a Socratic Study Buddy using Gemma 4. Unlike previous open models, Gemma 4’s high intelligence-per-parameter allows it to maintain complex pedagogical personas without "breaking character."
We will walk through the full developer lifecycle: from local setup in LM Studio to implementing multimodal vision inputs for analyzing textbook diagrams. We will also demonstrate how to bridge the gap between text and visuals by using Gemma 4 to generate real-time Mermaid.js logic flowcharts. Whether you are a student, educator, or developer, you will learn how to turn a raw model into a private, offline, and truly intelligent mentor. |
| 🌟 Speaker Tarun Jain : Run On-device AI Applications without internet using Qdrant Edge and Google LiteRT | This session introduces Qdrant Edge, an in-process vector search engine for embedded and resource-constrained devices, paired with Google's LiteRT for running a language model like Gemma-4 locally.
Together they let you build retrieval and inference pipelines that run entirely on-device. With no internet connection required. |
| 🌟 Speaker Farah Hammami: Responsible AI with Gemma: Build, Test, Break, Improve | Building an AI application is only the first step,how do we know it behaves reliably when users push it beyond the happy path?
In this hands-on workshop, participants will use Gemma to build a small AI-powered application and explore responsible AI through experimentation. Rather than treating responsible AI as a checklist, we’ll follow an iterative engineering approach: build, test, break, and improve.
Participants will challenge their application with real-world edge cases, including hallucinations, ambiguous requests, prompt injection, and questions outside the model’s knowledge. We’ll then analyze why these failures happen and implement practical techniques to make the application more reliable, grounded, and robust.
By the end of the workshop, participants will have hands-on experience building with Gemma, evaluating AI behavior, identifying weaknesses, and improving an AI application through iterative testing. |
| 🌟 Speaker Gabriel Preda: Fine-tune Gemma 4 models with Keras | Discover how to customize and optimize Google's latest Gemma 4 open models for domain-specific use cases using the modern Keras 3 ecosystem and JAX. This session provides a practical introduction to efficient fine-tuning techniques, covering the end-to-end workflow from data preparation and model adaptation to training, evaluation, and deployment.
Participants will learn how Keras simplifies large language model development across multiple backends, with a particular focus on JAX for high-performance, scalable training. Through hands-on examples, we will explore parameter-efficient fine-tuning approaches, distributed training capabilities, and best practices for achieving production-ready results while minimizing computational costs.
Whether you are building enterprise AI applications, experimenting with custom language models, or looking to leverage the latest advances in open-weight models, this session will equip you with the knowledge and tools needed to fine-tune Gemma 4 models effectively using Keras and JAX. |
