[Oct 2025] AI Community — Activity Highlights and Achievements
We love sharing the accomplishments of the Google AI communities over the month. We appreciate all the hard work and dedication of our community members. Without further ado, here are the key highlights!
Gemini CLI
Gemini CLI Extensions
Code Walkthrough and demos of the Nanobanana CLI extension by AI GDE Connie Leung (Hong Kong) demonstrates commands to call the tools of the Nanobanana MCP server to create and edit images, generate patterns, app icons, diagrams, and a story.
Unboxing and hands-on experience #geminicli externsions for #firebase by AI GDE Cyrus Wong (Hong Kong) demonstrates setting up a Firebase project and deploying a Gemini CLI extension. He covers installation, understanding files, building, linking, adding custom commands/GEMINI.md, and releasing the extension. He also shares how to use Gemini CLI for AI model interaction & integration, alongside debugging & troubleshooting.
A Developer’s Guide to Building Gemini CLI Extensions
A Developer’s Guide to Building Gemini CLI Extensions by Google Workspace GDE Kanshi Tanaike (Japan) provides a comprehensive walkthrough for developing Gemini CLI extensions, covering setup, configuration, local testing, and automating dependency management with GitHub Actions.
Gemini CLI Extensions for BigQuery: Natural Language Data Analysis by Google Workspace GDE Keisuke Oohashi (Japan) introduces Gemini CLI extensions for BigQuery. With a step-by-step guide on how to install and use the extensions, it shares how to set up prerequisites and environment variables.
Gemini CLI
Gemini CLI Commands website by AI GDE Sascha Heyer (Germany) is a collection of commands contributed by the community for Gemini CLI. You can find commands for various workflows, including UI implementation, security audits, backend development, code preparation, GitHub commits, and etc.
https://medium.com/media/2c1a1616e2d38d71153088521e32afdb/href
[👏152+] Build an AI That Connects to Your Database: MCP Toolbox for Databases Course (YouTube playlist) by Google Workspace GDE Aryan Irani (India) shares how to build AI agents that connect to databases using ADK, BigQuery, and the MCP Toolbox to address the Data Gap. The article guides how to enable access to real-time data for better responses, covering fundamentals and building advanced data agents for tasks like inventory management and sales analysis.MCP Development with the Google Cloud Rust SDK and Gemini CLI by AI GDE William McLean (US) leverages Gemini CLI to add MCP support for deploying AI apps built in Rust. He also extends the official Google Cloud Rust SDK to provide API call information over an MCP connection, validating that a compiled language like Rust can be used for AI software development. He also shared MCP Development in Rust with Cloud Run Deployment detailing how to develop and deploy a Rust-based MCP server to Google Cloud Run.
JAX
Snake game in JAX by AI GDE Victor Ashioya (Kenya) focuses on implementing JAX using PPO within a gamified environment to help explain how reinforcement learning works, following the original PPO algorithm implementation.
Build with AI: Fine-Tuning and Deploying Gemma Models with Keras & JAX by AI GDE Taha Bouhsine (US) was a live coding workshop guiding fine-tuning and deploying Gemma using Keras/JAX. He covered setup, weight loading, fine-tuning strategies (full training, parameter-efficient tuning, LoRA), JAX acceleration, and exporting models for CPU, GPU, and TPU deployment.
Multi-Modal Fusion: A Technical Deep-Dive on How Vision using JAX by AI GDE David Cardozo (Canada) took a technical deep-dive into the evolution, architecture, training strategies that make this possible, and how to implement them efficiently in JAX. It included live-coding of a simplified vision-language model to demonstrate JAX’s strengths in JIT compilation, automatic differentiation, and vectorization.
ADK • A2A • AP2
#AP2 LINE Bot Integration with Google Automated Payment using Agent Payments Protocol (repository | video) by AI GDE Evan Lin (Taiwan) discusses the integration of AP2 with a LINE bot using Gemini. He detailed the benefits for e-commerce integration, implementation of an enterprise-grade LINE bot architecture with enhanced agents and fault tolerance.
#AP2 [👏74+] Agent Payments Protocol (AP2) for Retailers by AI GDE Artem Nikulchenko (Switzerland) discusses and explores AP2 and its potential impact on retailers and the future of e-commerce by possibly replacing traditional sites with UI-less agents.
https://medium.com/media/0dc32c8df88775524344bed9b774b98c/href
#AP2 Episode 255 — Agonizing About Agent-to-Agent by AI GDE Allen Firstenberg (US) and AI GDE Noble Ackerson (US) dived into A2A and its importance, comparing it with MCP and exploring their fundamental differences. They also discussed the new AP2 and its potential to revolutionize AI agent interactions and transactions.
Gemini
Gemini 2.5 Computer Use Guide With Demo Project: Build a Job Search Agent by AI GDE Aashi Dutt (India) shares how to build an AI-powered job search agent using Gemini 2.5 Computer Use. The hands-on tutorial covers browser automation using Playwright and Streamlit without needing APIs.
Gemini 2.5 Computer Use: The AI That Uses Your Browser for You by AI GDE Carlos Alarcon (Colombia) gives a guide to Gemini 2.5 Computer Use, designed to autonomously control your browser. He demonstrates its ability to automate web work through practical examples like searching sites, solving captchas, and organizing tasks. Also explained the risks, security best practices, and how to experiment with the tool.
Building an AI Photo Editor with Angular, Firebase AI Logic, and Nano Banana (repository) by AI GDE Connie Leung (Hong Kong) ported Patrick Loeber’s Gemini 2.5 Flash Image examples to Angular and Firebase AI Logic, adding custom prompts for image generation from uploads. The key technologies used include Angular 20, Gemini 2.5 Flash Image, Firebase AI Logic, and Tailwind CSS.
Revolutionize Your Workflow with Gemini Code Assist (video) by AI GDE Dimitre Oliveira (Brazil) covers setting up and using Gemini Code Assist to improve coding workflow, including its GitHub and IDE plugin integrations and Agent Mode for complex tasks.
RAG with Gemini Codelab by AI GDE Jay Thakkar (India) is a codelab for the RAG implementation using Gemini to build intelligent chatbots, featuring document ingestion, vector search, and contextual query answering for enhanced accuracy.
Build a Custom Connector for Gemini Enterprise by AI GDE Sascha Heyer (Germany) explains how to integrate a system into Gemini Enterprise using a custom connector; and details the process for building a connector that ingests local files (Markdown, CSV, TXT) into Discovery Engine. He showed how to enforce fine-grained access control, and keep data current via live updates focusing on the Discovery Engine API.
Mastering Gemini Integration: Structured JSON for Your AI Applications (slides | blog post) by AI GDE Saverio Terracciano (UK) shared Gemini’s native support for structured output, which enables predictable data integration into robust applications using concise JSON schemas within AI Studio for defining, testing, and generating API code.
Long term Memory layer for Agents using Gemini and Mem0ai (slides | Colab Notebook) by AI GDE Tarun R Jain (India) guided how to build memory-aware agentic workflows using Gemini and Mem0ai to create AI agents with persistent memory, enabling them to learn and provide personalized experiences.
AI GDE Sascha Heyer (Germany) shared a series of articles on Exploring Google’s Multimodal Live API: Code & Articles. It provides practical guides into various aspects of Multimodal Live API along with how to leverage the powerful tools in your own projects. You can start here with the first article, A Developers Guide to Googles Multimodal Live API.
Gemma
https://medium.com/media/4a2dd9874e214ee923de4d4e92f55481/href
Introduction to Gemma by AI GDE Daniel Gwerzman (UK) introduces Gemma explaining its function, how it differs from Gemini, how to deploy and test it, and the benefits of SLMs over LLMs.
Develop an on-device RAG system powered by Gemma models (demo video) by AI GDE Georgios Soloupis (Greece) provides a step-by-step walkthrough for loading a PDF file, extracting and chunking its text, performing similarity matching, and using Gemma 3 to generate context-aware answers to user queries about the document.
Clinical Reasoning with MedGemma (Kaggle notebook | presentation note) by AI GDE Victor Ashioya (Kenya) explored the evolution of AI-assisted clinical reasoning, contrasting a general-purpose generative model approach with the enhanced capabilities of MedGemma. This talk showcased an upgraded workflow for robust healthcare AI applications.
PRs to Gemma
- Implementation of VaultGemma Fine Tuning with Differential Privacy and Inference #244 by AI GDE Rubens Zimbres (Brazil) added a complete pipeline for privacy-preserving fine-tuning and inference of VaultGemma 1B on medical data using LoRA and differential privacy.
- Custom Function Calling Router for Gemma-3 by AI GDE Tarun R Jain (India) introduced a custom function calling for Gemma 3. The model doesn’t need native tool tags, so he instructed it how to format a tool call using a structured prompt with delimiters. He also made a GitHub PR and it has been merged to the Gemma cookbook.
Gemma Day by GDG Ahlen
https://medium.com/media/3e4dabc42c603c294a5d0858d65b7825/href
DevFest Ahlen & Heilbronn 2025 — Gemma day by GDE Ahlen took a deep dive into Gemma in various aspects:
- Local Agentic RAG using Gemma and Qdrant (slides | Colab notebook) by AI GDE Tarun Jain (India) demonstrated how to build a completely local agentic RAG without any external APIs using Gemma 3n and Qdrant.
- Unleashing the Power of Gemma: From Cloud to Edge by AI GDE Vasudev Maduri (UK) explored the capabilities of Gemma models, providing practical insights, demos, and code examples for various environments such as Kaggle, Hugging Face, mobile, local machines, and Vertex AI. The session teaches how to customize Gemma for diverse tasks and harness its full potential.
Colab
Context wise ReRanker (Colab notebook) by AI GDE Mohamed Berrimi (France) guides on using DSPy and Gemini 2.5-flash to create a context-aware reranker for Retrieval and RAG applications, enhancing the relevance of retrieved documents by incorporating conversational context.
Keras
[SmolLM3] Add Backbone, CausalLM + Converter for HuggingFace Weights by AI GDE David Landup (Japan) introduces the SmolLM3 model to KerasHub, along with its backbone, causal language model, preprocessor, tokenizer, and associated layers.
AI GDE Harshal Janjani (UAE) contributed to KerasHub by adding MobileNetV5 and D-FINE to the library.
Kaggle
https://medium.com/media/a6f5b780beb29d969496711e15626778/href
Kaggle Community Olympiad — HACK4EARTH Green AI (video) hosted by Green Reliable Software Budapest successfully concluded, with 44 teams participating. The Demo Day showcased the 3 real-world solutions that tackle a greener impact.
AI GDE Dimitre Oliveira (Brazil)’s guide on how to Create a Remote LLM Server Using Kaggle Notebooks and Ollama were featured on Google for Developers’s X channel. It went viral with 857+🩷 and 177+🔄.
Vertex AI
Google ADK + Vertex AI Live API
[👏83+] Google ADK + Vertex AI Live API: Building Streaming Experiences by AI GDE Sascha Heyer (Germany) explores how to integrate Vertex AI Live API with ADK to create custom streaming experiences. It provides a guide for server-side integration and covers ADK initialization, media handling, and receiving responses.Building a Multi-Stage RAG Pipeline with Google ADK and Vertex AI by AI GDE Gabriel Preda (Romania) introduces a three-stage RAG pipeline (retrieval -> analysis -> answer) implemented using ADK, Vertex AI, and Gemini 2.5 models. The system uses a Vertex AI RAG corpus populated with EU AI Act documents. Agents are orchestrated using ADK’s SequentialAgent, including Retriever, Analyzer, and Final Answer agents.
Deploy an Agentic AI Application on GKE with ADK and Vertex AI by Biswanath Giri (India) showcased how Vertex AI enabled the end-to-end containerization and GKE deployment of the ADK app, ‘capital agent,’ and provided core access to the Gemini.
AI Research
An overview of DeepMind’s AlphaEvolve by AI GDE Grigory Sapunov (UK) introduces AlphaEvolve, a coding agent that uses LLMs to develop and evolve algorithms for user-defined problems. He detailed how it works and how it improves performance metrics based on a user-provided evaluation function. In addition, he shared an AI-powered version of a review of the same paper, created by an automated multi-agentic system for paper reviews.
https://medium.com/media/455e7ef61f2cd67f485a76351aae40d5/href
Neural Assets and World Models (slides) by AI GDE Martin Andrews (Singapore) explored the mechanics of AI simulation by examining Image Models (e.g., DallE-v3, Qwen) and World Models (e.g., MineRL, Dreamer-v4, GENIE). This talk emphasized the importance of automatic captioning and explanation for effectively training large generative models and concluded with a demonstration of the TinyWorlds model series, showcasing a custom Colab notebook he had developed for interactive use.
https://medium.com/media/49c9a9ca381a023e03cd0a24555bd159/href
ML Nashik successfully achieved the goal of AI Paper Reading Clubs (playlist) covering major AI topics (ROS-LLM, Attention Is All You Need, RAG-MCP, Agentic AI) and encouraging the participants to submit their projects for conference publication.
[Oct 2025] AI Community — Activity Highlights and Achievements was originally published in Google Developer Experts on Medium, where people are continuing the conversation by highlighting and responding to this story.
