NODE - AI Tinkerers São Paulo - Google Deepmind Hackathon
AI Tinkerers - São Paulo
Hackathon Showcase

NODE

NODE is a decentralized, offline-first mesh network using on-device AI to provide civilians with real-time resource mapping and safer navigation.

4 members Watch Demo

NODE – Life Support Assistant

NODE is an open source software built to be a survival companion in extreme danger situations — conflict zones, bombings, shootings, building collapses, and any scenario where every second matters and internet connectivity cannot be guaranteed.

The project’s core focus is simple: put alongside any person, anywhere, an assistant that speaks, listens, and sees — and that works 100% offline, directly on the device, without relying on the cloud, external servers, or an internet connection. This is what we call Edge AI: all intelligence runs locally, on the hardware itself.

In practice, the user can speak to NODE through the phone’s microphone and receive guidance in voice and text on how to act — where to take shelter, how to help an injured person, how to identify a threat. They can also point the camera at the surrounding environment and NODE analyzes the image in real time, identifying risks, escape routes, and shelter structures. The device’s geolocation is automatically used to further contextualize the responses.

Under the hood, the assistant uses Google’s multimodal Gemma 4 model to process both text and images, OpenAI’s Whisper to transcribe the user’s speech without sending anything outside the device, and native speech synthesis to respond in audio — all packaged in a single Docker container, ready to run with just one command.

The assistant’s behavior was carefully tuned: short, direct, and calm responses — like a trained person standing right by your side, with your life as the absolute priority.

Not applicable.

Docker — I used Docker to package the entire stack into a single container: Ollama FastAPI (Python) — I used FastAPI as the backend framework Gemma 4 (Google) — I used Google's gemma4:e2b model running locally via Ollama as the brain of the assistant. It's responsible for generating text responses and analyzing images of the environment — identifying threats Ollama — I used Ollama to serve and manage the Gemma 4 model locally. It was essential for running a multimodal LLM completely offline Products & Tools Used Python dependencies Whisper (OpenAI) — I used Whisper for offline speech recognition. The user speaks all without any internet connection. and Whisper transcribes it to text directly on the device and on Linux espeak-ng — both work completely offline. and return the model's responses. and shelter structures and the application. This way anyone can run NODE with a single command escape routes espeak-ng / say (macOS) — I used these speech synthesizers to convert the text generated by the model into audio. On macOS I used the native say command with the Luciana voice ffmpeg — I used ffmpeg to process the audio files received from the browser (in WebM and OGG formats) before passing them to Whisper for transcription. like having an AI server running on the machine itself. photos from the camera responsible for exposing the API endpoints that receive audio from the microphone the audio is recorded in the browser the model without needing to configure anything manually. without sending anything to any external API.

GitHub Repository

Summarizing URL...

Pitch Deck

Summarizing URL...