This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
I built Touch Grass, an entirely offline, privacy-first route generator for cyclists, runners, and hikers. It is designed to get people off their screens and into the real world by instantly generating dynamic, locally-aware navigation plans that don't require an active internet connection or a pricey cloud subscription. The application features a Python FastAPI backend communicating with a local Gemma 2B model (via Ollama) and a Next.js frontend with a high-contrast UI specifically optimized for outdoor visibility in bright sunlight
Who is it for?
It is for athletes, casual walkers, and outdoor enthusiasts who want to explore their local neighborhoods without broadcasting their exact starting location coordinates to a third-party cloud server.
Demo
Code
How I Built It
I built Touch Grass using a modern, local-first stack designed for offline outdoor endurance routing:
Local Inference & Model: Powered by Gemma 2B, running locally on macOS via Ollama for instant offline response times without cloud latency.
Backend: Built with Python FastAPI, which handles custom pacing calculations and strict spatial guardrails. Instead of trusting the lightweight model with raw math, Python calculates exact kilometer markers and feeds a strict structural template to Gemma, eliminating hallucinations or geographic teleportation.
Frontend: Developed using Next.js (TypeScript & Tailwind CSS) with a high-contrast, mobile-first design optimized for outdoor readability under bright sunlight. It features an interactive preparation checklist and a timeline-based navigation view.
Why Does Open Innovation Matter?
Open innovation and open-weight models like Gemma 2B made this project possible in three critical ways:
Zero-Connectivity Resilience: Outdoor endurance activities (trail running, hiking, and long-distance cycling) frequently take place in cellular dead zones, forests, or national parks. A closed, cloud-dependent API would fail entirely in these environments; a local open-source model runs anywhere.
Absolute Privacy: Route planning inherently requires sharing sensitive starting coordinates (like a user's home or current location). Proprietary cloud services harvest and store this telemetry, whereas local inference guarantees that private location data never leaves the user's hardware.
Zero Cost Experimentation: Endurance athletes iterate through dozens of route variations during a training block. Open-weight models allow unlimited free planning without incurring per-token API charges.
Prize Categories
Best Use of Gemma
Top comments (0)