Developer Somnath Das has submitted TouchGrass AI for Week 1 of the Hacktoberfest Open-Source AI Challenge, offering a pragmatic alternative to cloud-dependent wellness apps. The project utilizes Google's open-weight Gemma 4 model running locally via Ollama to generate personalized, phone-free grounding rituals. By keeping inference on the user's machine, it eliminates the privacy paradox often found in digital detox tools that harvest mood and location data.
Architecture and Offline Autonomy
The stack is built for simplicity and local execution, featuring a FastAPI backend that communicates with Ollama's REST API at localhost:11434. The frontend is a React 19 and TypeScript application styled with an earthy aesthetic, including a procedural breeze synthesizer powered by the Web Audio API. This ensures zero external audio downloads and complete functionality without an internet connection, making it ideal for remote workspaces or hiking trails where 5G is spotty.
Privacy and Cost Efficiency
Unlike subscription-based wellness platforms that lock features behind paywalls or rate limits, TouchGrass AI is MIT-licensed and relies on open weights. The system uses a strict temperature of 0.3 to guide Gemma 4 in returning structured JSON schemas for 3-step rituals. If the model is still loading, a heuristic fallback ensures users can still generate a plan, prioritizing immediate utility over perfect AI generation.
Key Takeaways
- Runs entirely offline using Ollama and Gemma 4, ensuring data sovereignty for sensitive mental health inputs.
- Features a procedural Web Audio API synthesizer for ambient sound, removing the need for heavy asset downloads.
- Built with a modern stack including FastAPI, React 19, and Vite, with a one-click launcher for local setup.
The Bottom Line
This is the kind of pragmatic AI tool we actually need: it solves a real problem without demanding your data or your credit card. It proves that local inference can power meaningful, human-centric applications beyond just chatbots.