[Apr 2025] AI Community — Activity Highlights and Achievements

[Apr 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!

AI GDEs at Cloud Next 2025

https://medium.com/media/cf9c36132be770e911c275ed6df04158/href

AI Madness at Google Next ‘25! Gemini, Veo 2, Agents, and More 🤯 by AI GDE Carlos Alarcon (Colombia) is an analysis video of key AI and Cloud announcements from Cloud Next 2025 in Las Vegas. Carlos highlights the new Gemini and multimodal models focused on content creation, new tools and agent frameworks, etc.

Several AI GDEs participated in the event as tech talk speakers, including Henry Ruiz (US), Margaret Maynard-Reid (US), Muhammad Farooq (US), Omotayo Olasunkanmi Aina (UK), Rubens Zimbres (Brazil), Sascha Heyer (Germany), Simon Margolis (US), Tomasz Porozynski (Poland).

Community Highlights

Build with AI: Agent Development Kit Hands-on Workshop by Deep Tech Stars was a detailed walk-through of how to use ADK and a deep dive into the concept of MCP. This session included a hands-on to build a simple agentic solution using ADK.

Build with AI — Team Technojam by Team TechnoJam covered practical uses of AI tools and techniques for solving real-world problems. The event included sessions on architecting autonomous solutions using agentic AI, and discussing MCP, A2A Protocol, ADK, etc.

https://medium.com/media/9366b7c0bae198611f0e9cccc1582813/href

Automating Patient Intake in Hospitals : Building multiagent app with ADK by Alagu Prakalya at Build with AI Mysuru event shared experience on developing an automated patient intake app using Gemini 2.0 Flash and ADK.

Build with AI Series — Chapter Two: The Rise of GenAI by TFUG Jaipur explored key concepts for building GenAI applications, along with AI agents, the A2A protocol, and ADK.

Build with AI — Pune by AI Pune featured two hands-on sessions: 1) Unleashing the power of Troubleshooting using K8sGPT on GKE, and 2) Build a Gemini-Powered YouTube Summarizer.

Build with AI: Hands-on Workshop — Gemini as the Judge for RAG Evaluation by AI Surat Community was to understand how to leverage Gemini to enhance RAG evaluation. Attendees learned a step-by-step framework for generating balanced evaluation datasets and gained insights into quantifying RAG system performance.

AI/ML Community Bolivia hosted a Build with AI and AI GDE Henry Ruiz (US) gave a talk, Building Multimodal Agents using Gemini. The session explored the potential of multimodal agents and how to leverage Gemini’s capabilities to build smarter and more interactive AI solutions.

Product Highlights

Agent Development Kit (ADK)

[253+ 👏] Agent Development Kit: Enhancing Multi-Agents Systems with A2A protocol and MCP server (repository) by AI GDE Rubens Zimbres (Brazil) explains how he developed a multi-agent system that uses ADK, A2A, and MCP. This multi-agent system enhances security by systematically evaluating user inputs for threats like SQL injection and XSS using regex and Model Armor. Rubens shows how a structured integration of ADK, A2A, and MCP collectively ensure robust, secure and scalable AI solutions.

[119+ 👏] Model Context Protocol (MCP) x Gemini — Deep Dive with code and explanation by AI GDE Punsiri Boonyakiat (Thailand) introduces MCP and the core concepts behind AI Agents, the ReAct framework, how Gemini uses Function Calling, and etc. The article also explains how to get started with the MCP SDK in Python, with practical examples for anyone who wants to experiment and build smarter AI solutions.

https://medium.com/media/ee50211c33c00a12b37b785b47471921/href

Topic Tonic: Crafting Curated Content with AI Agents (video | Kaggle Notebook) by AI GDE Dimitre Oliveira (Brazil) is an application that leverages the power of ADK and Gemini to automatically research, draft, critique, revise, and even narrate informational stories on any topic.

Google’s Agent Development Kit (ADK): A Guide With Demo Project (repository) by AI GDE Aashi Dutt (India) walks you through how to build a multi-agent, AI-powered travel assistant using ADK and the A2A protocol.

[ Google ADK ] Google tells everyone how to make LLM “start a family and build a career” by AI GDE Liu Yu-Wei (Taiwan) introduces ADK as a tool to develop AI agents and explains how Google thinks about the production method of AI agents, as well as the appearance of the large AI agents architecture.

Build with AI: Brand Monitoring Challenge by AI GDE Gabriel Cassimiro (Brazil) is a repository containing an exercise that encourages participants to attempt to use the ADK (optionally MCP) to build an end-to-end agent. The agent ingests and analyzes text from various online sources to produce a report detailing current public perception about a brand and potential areas of concern.

[ Open Source Project ] Google ADK Server Example by AI GDE Liu Yu-Wei (Taiwan) is an repository aims to simplify the process of deploying ADK agents in local or cloud environments, packaging the original example into a ready-to-run Docker image and exposing a RESTful API using FastAPI.

Build a ChatBot with Google ADK (Agent SDK) (repository) by AI GDE Evan Lin (Taiwan) shows you how to create a simple LINE Bot function as a starting project for MCP and other functions. Previously he also built the OpenAI Agents SDK version of the LINE Bot example.

https://medium.com/media/31141905d7751ff9f353fbaadd83fcd1/href

Google’s NEW Agent2Agent Protocol by AI GDE Sam Witteveen (Singapore) is a video introducing the A2A protocol, what it can do, how it will be in the future.

Gemini

Gemini 2.5 Pro: The BEST AI for Vibe Coding? 🤯 I Tested It (Code and Demos) by AI GDE Carlos Alarcon (Colombia) is a video showcasing Gemini 2.5 Pro’s reasoning and coding capabilities. Carlos demonstrates various tasks like creating a game app, extracting data from a PDF, and building a complex CRM system.

Gemini 2.5 Pro for Audio Transcription & Gemini 2.5 Pro for YouTube Analysis by AI GDE Sam Witteveen (Singapore) are videos introducing Gemini 2.5 Pro models capabilities for creating audio transcriptions & analyses (including audio diarization), and video analysis, respectively.

Image Segmentation using Gemini 2.5 (Colab Notebook) by AI GDE Nitin Tiwari (India) is a demo that shows how Gemini 2.5 models have improved spatial understanding, including segmentation tasks. He’s built a notebook that demonstrates how to segment any object using natural language, for both images and videos.

https://medium.com/media/2cb5446603e9f74ca55b404ba12a3e0b/href

Gemini 2.5 Flash — Hybrid Reasoning on Demand by AI GDE Muhammad Farooq (US) is a video exploring the Gemini 2.5 Flash, the first hybrid reasoning model that allows you toggle “thinking mode” on or off. Muhammad breaks down its cost‑performance ratio, 1M‑token context window, and how it stacks up against GPT‑4 Mini and Claude Sonnet.

Gemini Deep Research — Do Search the right way by AI GDE Tarun R Jain (India) introduces the new features, Deep Research in Gemini. It enables users to automate complex research tasks in depth and quickly. Tarun shares how this feature can be used for a variety of purposes for researchers, students/startups, non-technical, content creators.

Google’s AI Studio in 22 Minutes by AI GDE Muhammad Farooq (US) showcases the redesign of AI studio exploring new features including video generation and the various new models, hyper-parameters, and safety settings.

GenAI Usecase — Compare Product Prices via Gemini Grounding Search + Vertex AI Search on LINE Chatbot by AI GDE Punsiri Boonyakiat (Thailand) compares performance and prices of Gemini Grounding Search and RAG with Vertex AI Search for the task of information retrieval/reasoning and understanding them from image/PDF source.

https://medium.com/media/9769392c0e58f159715996b60c1d6e1f/href

Gemini Browser Use by AI GDE Sam Witteveen (Singapore) explores Browser Use & Web UI to connect AI agents with websites and browsers to use Gemini models and their performances.

AI-Powered-Blog-Generator by AI GDE Juan Guillermo (Mexico) and AI GDE Henry Ruiz (US) is an application that utilizes AI agents to automatically generate complete blog posts from a provided topic. It uses Gemini 2.0 with grounding to generate content, and the AG2 to orchestrate the interaction between different agents.

Natural Language Models: Interacting with Computers Like Humans by AI GDE Vinicius F. Caridá (Brazil) at Itau Data Week was a two-day lecture for more than 100,000 people. Vinicius spoke about the concepts of NLP, the SOTA Gemini models, and how AI Studio can help developers in this evolution.

Hands-on AI Agent Service: Based on Google Gemini API and Google Gemma Model by AI GDE Liu Yu-Wei (Taiwan) at Build with AI Changhua was a hands-on workshop exploring Gemini API and Gemma in depth from theory to practice.

Structured Output with Gemini Models: Begging, Threatening, and JSON-ing by AI GDE Saverio Terracciano (UK) shares tips for getting Gemini to return a structured JSON output and how AI Studio makes it easy.

Gemini API and OpenAI Compatibility: Voice Transcription Development in Actionby AI GDE Jimmy Liao (Taiwan) discusses how developers can integrate Gemini models using OpenAI-compatible interfaces. This integration allows applications originally built for OpenAI’s API to utilize Gemini models with minimal code changes.

Gemma

[191+ 👏 ] Fine-Tuning Google’s Gemma-3–4B for Reasoning: How GRPO Turned a Good Model into a Brilliant Thinker (Colab notebook) by AI GDE Eric Risco (Andorra) shares how he finetuned Gemma-3–4B to master logical reasoning, step-by-step, without explicitly teaching it.

https://medium.com/media/ffd3a3d8bb11f5a431d8c319f37bf99d/href

Run Gemma3 Anywhere! Colab, Local & Google Cloud (Ollama, Docker, OpenWebUI Tutorial) by AI GDE Ilias Papachristos (Greece) shows three ways to run Gemma 3 in a various environment — at Colab, Cloud Run, on local setup with Ollama, Docker Desktop, OpenWebUI.

GenAI on Your Own Terms — Running Open LLMs with Ollama on Colab by AI GDE Tomasz Porozynski (Poland) is a Colab notebook that provides a practical, hands-on demonstration of how to 1) set up Ollama in a Colab environment, 2) use Colab’s free T4 GPU resources, 3) run open LLMs, specifically Gemma 3 4B by default. This notebook was originally created for a specific talk but now he shared it publicly to empower people to try experiments with open LLMs.

Multi-modality inferencing and fine-tuning with Gemma3 by AI GDE leo KURI (China) was a codelab at Build with AI Nanjing. This session introduced how to run multi-modality inference with Gemma 3 and Hugging Face, followed by a hands-on guide to fine-tuning Gemma 3 with LoRA for efficient model adaptation.

Optimizing LLM Search Strategies: Training Gemma with GRPO for Diverse Query Generation by AI GDE M. Yusuf Sarıgöz (Türkiye) was a talk at Google /dev/cloud day Istanbul. He delved into a reinforcement learning (RL) approach using GRPO to train Gemma for generating diverse yet relevant search queries. Yusuf covered the reward function design, training process, evaluation metrics, and the impact on retrieval performance, offering insights into how RL can refine LLM-driven search strategies.

Deep-js-research by AI GDE Eric Risco (Andorra) is an open source research tool that helps you research any topic and generate factual research papers. Eric used Gemma for content generation and DeepSeek for analysis and planning.

https://medium.com/media/110a3614e86f19e41a6de9d71beb5cf8/href

Gemma 3 27B IT Abliterated (social post) by AI GDE Maxime Labonne (UK) is a demo of abliteration with refusal direction removal in Gemma 3 models. Maxime comments that the uncensored version of Gemma 3 created a very high acceptance rate (>90%). He also shared updated AutoQuant, his project to quantize LLMs, to make the GGUF versions of Gemma 3 abliterated.

#ODML Gemma 3 running with Ollama on a Nvidia Jetson Orin Nano computerby AI GDE George Soloupis (Greece) delves into running Gemma 3 with Ollama on a Jetson Orin Nano device, a new developer kit from Nvidia, and consuming the result on Android. This combination delivers low-latency AI inference without the need for cloud connectivity, making it ideal for embedded, real-time scenarios when privacy is also a priority.

Firebase Studio & Genkit

https://medium.com/media/413d9b2d25f360a9e00f159380c39c05/href

Google has launched @Firebase Studio! Build your app with AI for FREE and no coding required 🔥🧠 by AI GDE Yecely Aridai Diaz (Mexico) is a video introducing Firebase Studio and its main features to build your own AI-powered app from the browser in minutes.

Firebase Genkit Demo: Creating Acceptance Criteria and Test Cases from User Stories by AI GDE Juan Guillermo (Mexico) is a repository that demonstrates how to use Firebase Genkit to build an AI-powered application that generates acceptance criteria and test cases from user stories.

App Dev with Google’s Project IDX (and no coding!) by AI GDE Simon Margolis (US) guides how to write an application from scratch via Firebase Studio (formerly Projex IDX) without typing any code.

Build a Star Wars themed Agentic Application with Genkit JS, Gemini 2 0 Flash and Imagen 3 by Angular GDE Connie Leung (Hong Kong) shows how Connie built a Star Wars-themed agentic application using Genkit JS, Gemini 2 0 Flash, and Imagen 3. The app generates 500-word stories and poster images based on each character. Connie also conducted a Build with AI workshop on the same topic, guiding attendees on how to use Firebase Studio to create the app.

AI GDE Xavier Portilla Edo (Spain) contributed to the community plugin Firebase Genkit <> Mistral AI Plugin by updating it to the latest SDK and adding support for Mistral OCR.

JAX

The Art of Controlled Randomness: A Deep Dive into Sampling techniques in LLMs

The Art of Controlled Randomness by AI GDE Keshan Sodimana (Sri Lanka) explores advanced sampling strategies. The article explains the intuition and implementation behind key techniques — Temperature Scaling, Top-K Filtering, Top-P (Nucleus) Sampling, and Min-P Filtering — showing how each manipulates model outputs to balance diversity and reliability. With practical JAX code snippets and examples, it guides you through orchestrating these methods for fine-grained control over AI-generated language.

Cloud & VertexAI

Veo2 and Media Studio on Vertex AI — how to get started (console + colab) (Colab Notebook) by AI GDE Tomasz Porozynski (Poland) shows how easily you can generate images with Imagen3, and then use the images to produce video clips with Veo2.

Fashion with Imagen 3 and Veo 2: Part 1 (blog post) by AI GDE Margaret Maynard-Reid (US) is a tutorial focusing on Imagen 3 with a bit of Veo 2. Margaret shows how you can easily generate/edit visually consistent product photos using Vertex AI.

Google Veo-2 — The Best AI Video is Now Available to Everyone! by AI GDE Muhammad Farooq (US) introduces Veo 2 and its features, pricing, and how to use it.

Getting Started with Ray on Google Cloud Platform by AI GDE Henry Ruiz (US) introduces the fundamentals of Ray, an open-source framework for scaling AI and Python app, and walks you through how to deploy and manage Ray clusters on GCP using Vertex AI. Henry empowers you to run scalable and efficient distributed workloads with minimal operational overhead.

https://medium.com/media/cd99efe72881bbcfadc8d0f4556129bf/href

Deploying Powerful Language Models with Ease: Leveraging DeepSeek and Model Garden on Vertex AI by AI GDE Juan Guillermo (Mexico) is an article, based on his recent video, explores how Model Garden on Vertex AI simplifies accessing and utilizing open-source models like DeepSeek. Juan delves into the features of the Model Garden, the characteristics of DeepSeek, the synergistic benefits of using them together, and the associated considerations.

Building an Intelligent Movie Search with Neo4j and Vertex AI (demo | repository) by AI GDE Siddhant Agarwal (India) explores how you can leverage Neo4j and Vertex AI to build an intelligent movie search system. By integrating vector embeddings into a graph database, you enable semantic search that understands movie descriptions beyond simple keyword matching.

AI Training Campaigns

AI Prayagraj hosted the “Kaggle Community Olympiad — Unmasking Fakes” competition along with a related training session for beginners, which included completing Kaggle Learn courses and fostering active peer support.

ML Community Talk Agadir organized four sessions (playlist) that guided beginners on how to join the Kaggle Community Olympiad competition and make a submission.

https://medium.com/media/3e739f8e77533e09fe02bcc198eba4a4/href

GDG Ahlen is rolling out multiple campaigns in Q1, including AI Study Jams (playlist), AI Math Clubs (playlist), and Build with AI series.

[Apr 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.