Sherlawn Holmes: A Detective Game That Sends You Outside

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

Grass Hunt 🌿 — an outdoor scavenger-hunt web app that literally forces players to touch grass.

The Game Master describes their real neighbourhood on a phone (a neem tree, a Hanuman temple, a chai stall…), and an open-weight Gemma 4 model instantly writes riddles in that area's own language — Hinglish, Hindi, Gujarati or English. Teams then race outside against a countdown timer to find the real thing and photograph it. A second Gemma call judges every photo: is it the right target, and is it genuinely outdoors (no sneaky photos of a TV screen or Google Images)? Points, retries, funny in-language commentary, a scoreboard, and a Gemma-written sports-commentator recap at the end.

It's for kids, families, housing societies, school camps — anyone whose idea of fun has been stuck behind a screen. The screen is only the starting line and the referee; the game itself happens under the sky.

Demo

🔗 LIVE: https://sherlawn-holmes.onrender.com — open it on any phone, no setup needed. (Free Render instance sleeps after ~15 min idle, so the first hit may take 30–60 s to wake up.)

Prefer it offline? Run locally in 4 steps (see README) and open http://<your-laptop-IP>:5000 on any phone on the same WiFi — or flip USE_OLLAMA=1 and the whole game runs on a local Gemma model with zero internet.

• Video demo: (link here)

Code

GitHub repo: https://github.com/shreychauhan02/Sherlawn-Holmes

Deliberately tiny — no database, no login, no framework soup:

• app.py — one Flask file: Gemma 4 REST calls (cloud or local Ollama), defensive JSON parsing, error-safe routes
• templates/index.html — the whole mobile-first game UI in one file (inline CSS + JS)

How I Built It

The brain of the game is Gemma 4 (gemma-4-26b-a4b-it), Google's open-weight model, used three ways:

1. Riddle writer — one call turns the Game Master's area description into N rhyming scavenger riddles in the chosen language and difficulty.
2. Photo judge — a vision call checks {correct, outdoors, comment} against each uploaded photo (client-side compressed to ≤800px).
3. Commentator — a final call writes the funny match recap.

Gemini's generateContent REST endpoint is hit with a single swappable MODEL constant, and because Gemma doesn't offer JSON mode, all instructions live in the user prompt and the parser digs the JSON out of the model's reasoning preamble with a retry. A USE_OLLAMA=1 flag switches the whole game to a fully offline local Gemma via Ollama — same code, same game, zero internet.

Why Does Open Innovation Matter?

A closed, fixed-pricing API would have killed this project twice over. Grass Hunt makes many small vision calls per game — one per photo per team — and open weights mean I can run a whole society's tournament on a laptop with ollama and no billing anxiety at all. It also means the game runs in a village ground with no internet, schools can self-host it, and anyone can swap in tomorrow's better open model by editing one line. The model is a replaceable part, not a paywall — that's what an open-weight ecosystem makes possible and a closed API doesn't.

My Agent Session

Built with Qoder (agentic IDE): Flask backend, single-file game UI, live API debugging sessions (model 500s, Gemma 4 reasoning-preamble JSON parsing) all verified end-to-end against the real model.

Prize Categories

• Open-weight AI / Gemma category
• "Touch Grass" real-world impact