[Dev 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 3 Flash⚡
https://medium.com/media/875373853c48b27a292377ebe15cf39b/href
[👀39.4k+] Gemini 3 Flash — Your Daily Workhorse Upgraded by AI GDE Sam Witteveen (Singapore) goes through the latest Gemini 3 Flash model, highlighting its increased intelligence, speed, and affordability.Gemini 3.0 Flash Model Introduction — A model with better cost-performance and suitable for producing Agent products is officially launched by AI GDE Yu-Wei Liu (Taiwan) summarizes key information of the Gemini 3 Flash Preview.
Gemini 3 Flash Tutorial: Build a UI Studio With Function Calling by AI GDE Aashi Dutt (India) shares how to use Gemini 3 Flash to create a UI Studio for assembling dashboards using tool calls, structured outputs, and rapid, knob-based iteration.
Creating a document classifier and OCR: How to use Gemini 3 Flash to simplify the process (repository) by AI GDE Juan Guillermo Gomez (Mexico) walks through a demo app built to solve the operational challenge of manual document processing in loan origination — a costly bottleneck prone to error — by showing how pragmatic design and new AI features enable the rapid delivery of a scalable, intelligent document processing MVP.
Antigravity in Action: Real-World Use Cases
Antigravity
Antigravity Rules & Workflows Guide: AI-Powered Development Standards by AI GDE Jimmy Liao (Taiwan) introduces two mechanisms for enforcing development standards in AI-powered IDEs. The guide uses the agy-starter project to demonstrate how to use and customize Rules and Workflows for consistent code quality, time savings, and knowledge transfer within a team.
How to Use Cerberus for Keras: A Complete Antigravity Debugging Tutorial by AI GDE Harshal Janjani (UAE) built a PoC extension for Antigravity that automatically tests your models across TF, JAX, and PyTorch, all with a single save.
Building Scalable Flutter Apps with Google Antigravity: From Idea to Product and Keras for Researchers: Rapid Prototyping of Neural Nets for AI Project by GDG Ahlen (Germany) each explore how Antigravity functions as a full-stack partner for scalable Flutter development and why Keras serves as the premier engine for AI research.
Antigravity with Colab
[👏107+] Leveraging TPUs in Colab — featuring Antigravity by AI GDE Glen Yu (Canada) is a follow-up to his previous work on Colab in VS Code. In this post, he demonstrates TPU performance for fine-tuning Gemma and explains how Antigravity helped optimize his TPU-based development workflow.
Running Colab on Antigravity by AI GDE William McLean (US) is a demo on using the Colab hosted notebook environment as a plugin directly in Antigravity. It shows how to generate and execute a complete notebook directly.
Antigravity Quota Extension: Real-time Monitoring of AI Model Usage by AI GDE Jimmy Liao (Taiwan) is a VSCode extension that displays Antigravity AI model quota usage in the status bar. It solves the problem of developers needing to manually check quota usage, which interrupts workflow. This extension provides real-time usage percentages, a countdown to quota reset, customizable model display, and manual refresh.
JAX/Flax
[👏302+] Utilizing LLM Embedding to Create a Classification Model for Sentiment Analysis using JAX and Keras by AI GDE Joan Santoso (Indonesia) explores using LLMs for SOTA embedding techniques. It demonstrates building a sentiment analysis DNN using Keras and JAX/Flax for robust model construction.https://medium.com/media/df4f9b5bb3a61825c4df299394c48fe4/href
Can mass ophthalmology screening be made affordable through open-source? by AI GDE Samuel Marks (US) shares a mission to combat blindness by combining low-cost hardware with SOTA AI. He shows how JAX and OpenXLA can optimize DL models for high-speed, medical screening, making ophthalmology tools accessible in resource-limited settings.
Building Scalable Models in Pure JAX by AI GDE Aakash Nain (India) is a guide to building production-ready ML models in pure JAX with minimal abstractions that scale to thousands of accelerators.
Learn Flax & JAX: Neural Network Training Guide by AI GDE Taha Bouhsine (US) is a learning resource for training neural networks with Flax and JAX.
Marin 32B Retro: Scaling and Stabilizing a Large Language Model by AI GDE David Hall (US) details the training process (on TPU v4), challenges, and solutions. It covers scaling from the 8B recipe, addressing loss instability with QK-Norm, and resolving data contamination and shuffling anomalies during cooldown.
Build a Transformer LLM with JAX from Scratch (slides) by AI GDE Lai Fong Leong (Malaysia) explores developing a Transformer model using JAX for the beginners and guides how to develop all the components from scratch.
DL Frameworks and Open Source Acceleration Libraries (slides) by AI GDE Joe Halabi (US) introduced the latest open source DL ops acceleration libraries and their potential usefulness in DL Frameworks such as XLA and JAX.
ADK
Deploy your ADK Agentic application on Cloud Run by AI GDE Gabriel Preda (Romania) walks you through how to deploy an ADK agentic application to Google Cloud Run as the next step after building and refactoring your SQL-based ADK agent.
Building a Store Search Agent with Google ADK and Maps Grounding Lite MCP by AI GDE Kazuki Hara (Japan) is a hands-on guide building a store search agent using ADK combined with Google Maps Grounding Lite MCP. The agents can easily access geospatial data such as place search, weather forecasts, and route calculations without dealing with complex API calls. Project: A demo of searching for recommended ramen shops within 10mins walk from Shibuya Station
AgentOps — The Blueprint for Enterprise AI Agent Engineering with Google Cloud, Gemini 3.0 and ADK by AI GDE Shreyak Gupta (US) shares a comprehensive methodology for building, deploying, and managing reliable, scalable, and governed AI agents within the Google Cloud ecosystem.
AI Crafting Living Worlds: Next-Gen NPCs with Gemini and Gemma AI (slides) by AI GDE Jai Campbell (UK) explored the creation of dynamic and responsive NPCs using Gemini/Gemma to adapt to player interactions, enhance gaming experience, and demonstrate features like state/memory management and agent-to-agent communication using ADK, GCP, and Firebase studio.
The Five Pillars of Calibrated Trust: Building Agentic Systems That Enterprises Actually Deploy (video) by AI GDE Noble Ackerson (US) and hosted by Machine Learning Lagos (Nigeria) demonstrates how to move beyond simple automation by implementing a trust-centered framework to build reliable autonomous systems for the enterprise.
Gemini CLI
Automating DevRel: How I Use Gemini CLI and Gemini 3 to Catch Bugs in My Blog Posts by AI GDE Connie Leung (Hong Kong) explored how to build an automated Technical Editor in the terminal using the Gemini CLI and Gemini 3 Pro Preview to streamline proofreading and content polishing.
[👏63+] Deploy MCP with Gemini CLI and Python on Google Cloud Run (repository) by AI GDE William McLean (US) demonstrates deploying an MCP with Gemini CLI and Python on Google Cloud Run. It involves testing locally with streaming HTTP transport before deploying to Google Cloud Run, streamlining the development process.Figma × Gemini CLI: How I Did It and Why You Should Too (Darija) by El Bachir Outidrarine (Morocco) shares how to generate accurate code and automate asset retrieval by leveraging the Figma MCP server to access exact design data.
Gemini
#Interactions API [👀51.9k+] The Gemini Interactions API by AI GDE Sam Witteveen (Singapore) looks at the new Gemini Interactions API, and how you can use it to build and do various tasks with not only Gemini models but also agents.
#Interactions API 🤖 A First Look at the Gemini Interactions API by AI GDE Jimmy Liao (Taiwan) explores its features like automatic state management and tool orchestration, as well as cost, data retention, and architectural considerations.
#Interactions API Goodbye to manual research: This is how Google’s Deep Research works by AI GDE Juan Guillermo Gomez (Mexico) explains Gemini’s Deep Research feature, detailing how the agent autonomously synthesizes complex information using web search and its own data, and analyzes the Interactions API to interact with Gemini models/agents.
Building a News and Information Assistant with Google Search Grounding API and Gemini 3.0 Pro by AI GDE Evan Lin (Taiwan) details the process of building a news and information assistant, highlighting its advantages like improved speed and quality, and provides solutions for common development issues with code examples for integration into a LINE Bot.
Gemini Live API Proactive, in Next.js and React Native Expo (repository) by AI GDE Felipe Lujan (Canada) explains how to implement Gemini’s Live API with proactive features, detailing implementation steps for both Next.js and React Native environments. Project: A live negotiation coach that helps the user get feedback while negotiating in person
Gemini for Library Maintenance (slides) by AI GDE Sayak Paul (India) discusses using Gemini or other models to assist with library maintenance tasks. The talk covered automated test suite design; leveraging LLMs for code refactoring ideas; aiding debugging; creating tooling to analyze code; and PR review feedback.
Building Multi-Model AI Agents: Model Selection, Reasoning, and Performance at Scale (slides) by AI GDE Kondwani Nyirenda (Zambia) focused on the shift from single-model AI to agentic intelligence, where a controller orchestrates multiple specialized models to solve complex real-world tasks. The talk covered model selection strategy; the golden rule of performance and optimization agent architecture, and real-world use cases.
The Latent Digital ‘Personality’ in Generative AI Models by AI GDE Ismael Chaile (Spain) explores the concept of ‘personality’ behind Gen AI models. It presents a preliminary experiment driven by curiosity, doing an initial basic research & development of whether these models can exhibit distinct characteristics. It also shares summaries of the Gemini models’ personality with generated images of them.
Nano Banana Pro 🍌
https://medium.com/media/ae80ba2b3d75ac8383c64332d22f4d64/href
🌍 Making Learning Fun and Accessible: How Gemini 3.0 Transforms Presentations for Global Education by AI GDE Cyrus Wong (Hong Kong) showcases his project, Gemini PowerPoint Sage, a 12-agent AI system powered by ADK and Gemini, which transforms static slides into engaging, multilingual video lessons. It automatically generates speaker notes & visuals, synthesizes audio, and creates videos in over 16 languages.
https://medium.com/media/3cee2e70be64def41f2cd9c75a9b5ab4/href
[👏51+] Fashion Moodboard with Gemini 3 & Nano Banana Pro by AI GDEs Margaret Maynard-Reid (US), Chansung Park (Korea), and Sayak Paul (India) introduces a visual tool that generates structured fashion moodboards from text prompts using Gemini, allowing designers to refine images with targeted edits.
Google Gemini Vision in FiftyOne by AI GDE Adonai Vera (Colombia) is a Colab notebook that demonstrates how to use the Gemini Vision Plugin with FiftyOne to analyze a real-world autonomous driving dataset, identify dataset issues, and use Gemini’s generative capabilities to improve data quality.
Use Gemini 3.0 Pro Image API to Create a PDF Text Optimization Tool by AI GDE Evan Lin (Taiwan) shares the creation of an automated tool using Gemini to enhance text clarity in PDF documents, specifically addressing blurry Chinese characters often generated by AI models.
Gemma
Gemma from Scratch: Mastering LLM Implementation with JAX and Flax
Your AI from Scratch: Building a Customised LLM (repository) by AI GDE Luca Massaron (Italy) outlines how to build a customized LLM (specifically referencing Gemma 270M) from scratch — starting from an empty Python file and ending with a fully functional, training-capable engine. He also shared a version of the guide using JAX/Flax, Gemma from Scratch: Mastering LLM Implementation with JAX and Flax (repository) with a focus on high-performance functional programming.
#FunctionGemma FunctionGemma — Function Calling at the Edge by AI GDE Sam Witteveen (Singapore) introduces the new model for on-device function calling.
[👏201+] Run Gemma3 Locally and Offline in R and Python Using Docker Model Runner by AI GDE Ifeanyi Idiaye (Nigeria) includes how to set up Docker Model Runner for local inferences and demonstrates how to interact with Gemma3 from within R and Python scripts.Cloud
Generating Audio-Driven Talking Face Animation from Single Image by AI GDE Muhammad Ghifary (Indonesia) presents an one-shot workflow for realistic audio-driven talking face animations. Utilizing Veo for dynamic movement and MuseTalk for high-fidelity lip-synchronization, it demonstrates that professional-grade digital humans are achievable with accessible commercial compute resources like NVIDIA L4 GPUs on Vertex AI Workbench.
Deep Dive into AI Agent Memory with Vertex AI Agent Engine Memory Bank by AI GDE Rio Kurihara (Japan) shared how to use Vertex AI Agent Engine Memory Bank for giving AI agents long-term, cross-session memory. It also explores practical applications such as pre-loading user history and customizing memory prioritization.
Building Location-Aware AI Applications with Google Maps Grounding API by AI GDE Evan Lin (Taiwan) discusses how to use the Maps Grounding API via Vertex AI for location-aware AI applications. It highlights the differences between Gemini Developer API and Vertex AI API, emphasizing that Maps Grounding is only supported through Vertex AI.
Getting Started with Google ADK by Building a Pokémon TCG Pocket AI Agent by AI GDE Kazuki Hara (Japan) is a tutorial on building AI agents with ADK, deploying them to Vertex AI Agent Engine, and integrating with Gemini Enterprise. Using a Pokémon TCG Pocket card search agent example, it walks you through the entire development-to-deployment workflow for beginner engineers.
Kaggle
Participations in Google DeepMind — Vibe Code with Gemini 3 Pro in AI Studio:
Lumina: Your Journey, Refined & Connected (repository) by AI GDE Chansung Park (Korea): An ‘AI Life Companion’ that supports connecting small habits to grand dreams. It records and analyzes daily life, and then generates a narrative with images in the final form of a blog post.
Feedforward — News at the AI Era (app trailer | video) by AI GDE Dimitre Oliveira (Brazil): A multimodal, TikTok-style news feed generated in real-time by Gemini 3 Pro, Search, and TTS.
*Additional blog post on [👏81+] How I “Vibe Coded” the Future of News with Gemini 3
AI Research
AI GDE Vinicius Caridá (Brazil) shared a review of Google Research’s proposal of Nested Learning: [👏50+] Nested Learning: A New Paradigm for Continual Machine Learning.
AI GDE Aakash Nain (India) shared an annotated paper and summary of On the Theoretical Limitations of Embedding-based Retrieval from Google DeepMind.
AI GDE Grigory Sapunov (UK) shared an article, The Transformer Zoo Revisited, that analyzes a recent paper, Encoder-Decoder or Decoder-Only? Revisiting Encoder-Decoder Large Language Model from Google DeepMind.
AI Training Campaigns Summary | 2H 2025
Throughout the second half of 2025, AI BDCs held 51 AI Study Jams, 18 AI Math Clubs, 21 AI Paper Reading Clubs, and 7 Kaggle Community Olympiad competitions. Here are the key event summary:
AI Math Clubs | 18 events by 9 communities
- GDG On Campus ISAMM (Tunisia) completed a year-long initiative for AI Math Clubs, resulting in 5 videos (playlist).
- Fatima Jannet (Bangladesh) also helped students master math concepts for AI (playlist).
AI Paper Reading Clubs | 21 events by 11 communities
- ML Nashik (India) hosted 4 paper review events (playlist) along with the paper writing training sessions aiming at conferences.
- TFUG Islamabad is rolling out AI Paper Reading Clubs (playlist) covering fundamental AI research papers.
Kaggle Community Olympiad | 7 competitions
Seven communities successfully achieved the goal of the Kaggle Community Olympiad 2025: AI Durg, ML Nashik, Machine Learning, AI, Deep Learning & NLP Community — Bangladesh, TFUG Bhubaneswar, The Coding Culture, Green Reliable Software Budapest, GDG on Campus IPEC and TFUG Hyderabad
*Tech talks and training sessions on submission guidelines (video) and on winning solution demos (video) were held alongside the competitions.
AI Study Jams | 52 events organized by 29 communities
Playlists from AI Community Lucknow, TFUG Islamabad, and GDG Ahlen.
[Dev 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.
