This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass Seekr is a live, map-based treasure hunt for real-world events. Organizers pin riddles and hidden locations on a map. Teams of up to 4 solve a riddle to unlock the next spot on their phones, walk there, and prove they arrived with a photo or a QR scan. A local AI writes hints (3 per team for the whole hunt) and checks that photos match the target. The first team to find the treasure wins, and the organizer watches every team move live. It’s for schools, campuses, festivals, team-building days and anyone running outdoor games.
Drive Link(Video): [https://drive.google.com/file/d/1-cqx2aITLNqumZaivHAWkZtjptTwAAg-/view?usp=sharing](https://drive.google.com/file/d/1-cqx2aITLNqumZaivHAWkZtjptTwAAg-/view?usp=sharing)
github: [https://github.com/73LIX/Seekr](https://github.com/73LIX/Seekr)
AI: Gemma (gemma4:e2b), an open-weight model, runs locally through Ollama. It writes the hints and compares each player’s photo with the organizer’s reference photo, scoring it out of 100. A score of 85 or more passes, and borderline scores go to the organizer for review.
Safeguards: a check stops hints from leaking the answer. If the AI is slow or offline, the hunt falls back to the organizer’s own hints.
App: plain Node.js with SQLite and no build step. Live updates use server-sent events, and the QR codes are generated and decoded by code I wrote. The pixel-art interface is plain HTML, CSS and JS.
Testing: the project has over 100 automated tests plus a full browser end-to-end run.
An open model made this possible without a cloud service: it runs on the organizer’s own laptop. That means no API fees, no usage limits, and players’ photos and locations never leave the venue. It also works on a field trip or at a festival with poor internet, and anyone can inspect, swap or improve the model. A closed API would have added per-photo costs, sent every player’s photo to a third party, and stopped working whenever the internet or the provider did.