![[sep-2025]-ai-community — activity-highlights-and-achievements](https://prodsens.live/wp-content/uploads/2025/10/40306-sep-2025-ai-community-activity-highlights-and-achievements-550x262.jpg)
[Sep 2025] AI Community — Activity Highlights and Achievements
Let’s explore highlights and accomplishments of the Google AI communities over the month. We appreciate all the hard work and dedication of our community members. Here are the key highlights!
Featured Story
A Practical Guide to Google’s ArrayRecord
AI GDE Minho Ryu (Korea) created a comprehensive “how-to” guide for ArrayRecord format, which is recommended by projects like MaxText for large-scale, distributed training. This guide provides practical steps and explanations to help developers, addressing the current lack of accessible documentation.
>> The guide / Project introduction / Github repository
Product highlights
Gemini CLI
AI GDE William McLean (US) shared three articles leveraging Gemini CLI and its power into the development process:
- [👏142+] Agent Development with Gemini CLI: demonstrates how to extend a Google codelab using the Gemini CLI for code explanation and generation, specifically for AI agents using the ADK and MCP.
- [👏205+] Code Generation with Gemini CLI, MCP Toolbox, and Firestore: a step-by-step guide on using Gemini CLI for code generation, combined with MCP Toolbox to analyze a Firestore Database.
- [👏55+] Refactoring Legacy APIs with Gemini CLI and MCP: demonstrates using Gemini CLI and Gemini 2.5 Pro for code refactoring to enable the MCP protocol with an existing Typescript codebase, using the Cymbal Super Store sample app.
Getting Started with Google ADK by AI GDE Vrijraj Singh (India) is a codelab on the fundamentals of ADK, providing the knowledge to begin building your own AI agents. Covered key concepts, setup, and practical examples to kickstart an AI journey.
Automate Your GitHub Workflow: Meet Your New AI Coding Partner by AI GDE Daniel Gwerzman (UK) is an overview of the Gemini CLI GitHub Action, a new AI coding tool that integrates directly into GitHub repositories. It automates issue triage, fixes bugs, and reviews pull requests.
JAX & TPU
The Winner is Quiet: Google Accelerates to AI Dominance with 7th Generation of TPUs and JAX/XLA Innovations by AI GDE Nguyen Khanh Linh (Vietnam) discusses Google’s advancements in AI hardware, specifically the 7th generation TPU known as Ironwood. The article highlights the TPU’s architecture, performance advantages, and integration with JAX and XLA for optimized AI workloads.
Intro to GenAI Processors hosted by Machine Learning, AI, Deep Learning & NLP Community — Bangladesh and presented by Victor Ashioya (Kenya), introduced TPU, its specialized architecture, and a paradigm shift in AI hardware design. It covered from the introduction in 2015 to the latest 7th generation, Ironwood, and how it consistently pushed the boundaries of what’s possible in ML acceleration.
[PR] Add Qwen3-[4B,8B,32B] on bonsai repository by AI GDE Minho Ryu (Korea) added support for the Qwen3 model series, specifically the 4B, 8B, and 32B parameter versions, to the jax-ml/bonsai library. He also Added Best-Fit Packing Algorithm and Optimization to the Grain library, a Python library for reading and processing data for training and evaluating JAX models.
[PR] AI GDE Rubens Zimbres (Brazil) submitted three PRs introducing three models — U-Net, EfficientNet, and VAE. The implementations used the flax.nnx API. Each PR includes the full model definitions, factory functions, and comprehensive example notebooks with validation.
– U-net Implemented in JAX PR #40
– Efficientnet Implemented in JAX PR #39
– Variational Autoencoder (VAE) Implemented in JAX PR #38
[Notebook] Proximal Policy Optimization (PRO) in JAX by Victor Ashioya (Kenya) is a top-to-bottom implementation of the PPO using JAX following the actor-critic architecture to understand the basics of policy gradient. Basically, training an RL agent to learn how to take actions in an environment to maximize rewards, using one of the most popular and effective RL algorithms.
[Notebook] Deep learning introduction using JAX (Colab notebook) by AI GDE Amel Sellami (Tunisia) was a workshop at Deep Learning Indaba, the annual meeting of the African ML/AI community. The session explored the fundamental concepts of deep learning and how to build/train models from scratch. Then it covered how to train larger models efficiently using JAX. She also led Introduction to LLM (Colab notebook) at the same event. This session aimed to introduce LLM to beginners, for understanding the idea behind Attention and why it is used; the fundamental building blocks of the Transformer architecture; and comparison tokenizers across different languages. This session also used JAX as a main framework
ADK
Building Intelligent Agents with Google ADK by Cloud GDE Amulya Bhatia (Germany) is a comprehensive ADK guide from set-up to deploying agents and security best practices for ADK agents. The book-style content of this 24+ chapters guide will help you get started with creating your own agent system.
Responsible AI and Agents on ADK (one pager) by AI GDE Ahirton Lopes (Brazil) was a talk at Google Cloud Summit Brazil. He presented his work on “Responsible & Trustworthy AI” and shared best practices, from governance and risk-mitigation frameworks to agent architectures powered by Gemini and ADK. A key highlight was presenting his role in the testing of OTTO — an embedded AI system built on Gemini and integrated into the new Volkswagen’s vehicle, with native AI, for the Brazilian market. He demonstrated how Gen AI can be embedded to connect cloud capabilities with the automotive ecosystem securely and scalably.
Agent Development Kit — ADK x MCP x A2A Workshops (slides) by AI GDE Punsiri Boonyakiat (Thailand) was a comprehensive workshop covering ADK basics, MCP integration, multi-agent design and A2A setup.
https://medium.com/media/b4f968904b327b91b97c639e15d3f6d1/href
AI Agent Development Kit (ADK): A Beginner’s Guide (slides | repository) & Agent2Agent (A2A) Protocol: A Primer by AI GDE Kuan Hoong Poo (Malaysia) were workshops covering ADK and A2A, about their core purpose, key features, and real-world use cases.
Deploying ADK to Agent Engine on Google Cloud with Next.js Integration (repository) by AI GDE Rodgers Ndocha (Kenya) covered production-ready full-stack deploy guide. Participants learned deploying from ADK to the Agent Engine on Google Cloud and integrating it seamlessly with a Next.js frontend.
Gemini
Vibe Coding Hackathon with Firebase Studio & Gemini by Deep Tech Stars showcased the winning apps that leveraged Firebase Studio and Gemini. The result apps are varied including agriculture disease diagnosis, chat security analysis, and brand asset generator.
Building a LINE File Backup Robot with Golang to Google Drive (repository) by AI GDE Evan Lin (Taiwan) details the creation of a LINE bot using Golang that automatically backs up multimedia files from LINE chats to Google Drive. The bot organizes files into folders by year/month and provides a simple file query function.
Gemma
[Kaggle] GeoSquire — AI Assistant for Geoscientists (repository) by AI GDE Felipe Lujan (Canada) is an Android app powered by Gemma 3n for geoscientists who do not have internet access during their field work. It helps geoscientists in mining, oil & gas and geology by classifying outcrops, rocks and identifying minerals to assist with geoscience-related tasks, data analysis, or modeling.
[Kaggle] InsightAid: Private AI Accessibility Assistant (repository) by AI GDE Tomasz Porozynski (Poland) is an AI accessibility assistant leveraging Gemma 3n. It focuses on providing private AI solutions for users with specific needs, demonstrating the potential of local AI models in enhancing user experience and accessibility.
Building a RAG System with PostgreSQL, pgvector, SQLAlchemy, and Gemma/Gemini Embeddings by AI GDE Henry Ruiz (US) covers generating a synthetic recipe dataset using Gemini, creating embeddings using both Gemini/Gemma models, storing these embeddings in a PostgreSQL database with the pgvector extension, and performing semantic retrieval using cosine similarity. It details the steps involved in setting up the database, creating tables and indexes, and populating the database with the generated data and embeddings.
Fine-tune Gemma-3–270M for Financial Sentiment Analysis (Kaggle notebook) by AI GDE Luca Massaron (Italy) is a hands-on tutorial & a result summary (notebook and article) detailing the fine-tuning Gemma 3 370M. Following up his previous work, Fine-Tuning Gemma 3 1B-IT for Financial Sentiment. Analysis: A Step-by-Step Guide, this project demonstrates that even the compact model can achieve a high accuracy highlighting its practical value for resource-efficient tasks.
EmbeddingGemma Introductions:
- [Post] Embedding Gemma running on device by AI GDE Georgios Soloupis (Greece)
- [Video] Embedding Gemma: On-Device RAG Made Easy by AI GDE Muhammad Farooq (US)
Vertex AI
New feature release for Agent Engine: Create A2A agents with ADK and Agent Engine using A2A support by Workspace GDE Keisuke Oohashi (Japan) discusses the A2A protocol support in Vertex AI Agent Engine. It covers creating server-side and client agents, deploying to Agent Engine, and using RemoteA2aAgent. It also highlights the benefits of native A2A support for building scalable multi-agent systems.
Leveraging Google Cloud for High-Performance ML Pipelines (video) by Machine Learning Lagos and AI GDE John Robert (Nigeria) shows how to set up an ML pipeline, starting with a basic Vertex AI workbench configuration and progressing to a complex design that integrates BigQuery for data and uses AutoML for training.
Shared notebooks
Data Analysis with Gemini in BigQuery Studio (Colab notebook) by AI GDE Juan Guillermo Gomez (Mexico) was a hands-on session learning how to use Colab Enterprise Python notebooks and BigQuery DataFrames for efficient data analysis. They explored how Gemini can generate code, streamlining the development of models like K-means clustering and the creation of visualizations.
AI GDE Arnaldo Gualberto (Brazil) shares a series of notebooks for machine learning basic algorithms:
– [Colab] Linear Regression, Multivariate Regression, Polynomial Regression, Logistic Regression, K-Nearest Neighbors
– [Kaggle versions] Linear Regression, Multivariate Regression, Polynomial Regression, Logistic Regression, K-Nearest Neighbors
GabrielPreda (Romania) shares three Kaggle notebooks for Data Science:
Kaggle: World Countries GDP 2020–2025, Medical Insurance Cost: EDA and Predictive Models, Crocodiles Species Around the World
Community Spotlight
https://medium.com/media/85adc7b32852cf7a721dad2b8f111379/href
GDE Connect Episode 1 | Global AI/ML Perspectives by Machine Learning Pakistan is a video focusing on global perspectives in AI/ML. Four AI GDEs — Muhammad Ahsan (Sweden), Kshitiz Rimal (Nepal), Mashhood Rastgar (Pakistan), and Xihan Li (China) — shared their thoughts on global AI trends, AI for developers, and AI for developing nations.
[Sep 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.
