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

What I Built

Walking Challenge AI is a small web app that turns an ordinary walk into a game.

You pick three things: how long you will walk (10 to 60 minutes), where you are (park, city, college or neighborhood), and how hard you want it (easy, medium or hard). The app then gives you a short list of real-world challenges, like "find 3 different types of trees" or "spend 2 minutes just observing the sky".

The screen is the shortest part of the experience. You spend about 20 seconds on your phone, put it in your pocket, and go outside. When you are back, you tick off what you completed, and a "Walk Complete!" screen closes the loop.

I built it for students like me who spend most of the day on screens and need a small, fun reason to step outside.

Demo

Live demo: https://walking-challenge-ai.vercel.app/

The first run downloads the model (about 1.2 GB), so use Chrome or Edge and give it a minute. After that it is cached by the browser.

The flow is simple: choose time, place and difficulty, get your challenges, walk, tick them off, finish.

Code

npm install
npm run dev

How I Built It

• Frontend: React, Vite and Tailwind CSS
• AI: an open-weight Qwen2.5 Instruct model running directly in the browser with Transformers.js (WebGPU when available, WASM otherwise)
• Safety net: a hand-written bank of challenges, so the app never breaks

The app builds a prompt from the walk time, place and difficulty and sends it to the model running locally in the browser. It then validates the output. If the model returns something unusable, the app falls back to the built-in challenges, so the walker always gets a working list.

What I learned: my first attempt used a 0.5B model, and it simply echoed my prompt back instead of following it. Small models are unpredictable when you ask for structured output. That is why the app checks the model's output before showing it, and why I moved to a larger open model. Making small-model output more reliable is the next thing I want to improve.

Why Does Open Innovation Matter?

• Runs locally, no server: after a one-time model download, everything happens in the browser. There is no API key, no backend and no cost per request.
• Privacy: the walker's time, place and activity never leave their device. That matters for an app about where people physically go.
• Swappable model: the model is a single line in the code. Anyone can replace it with a smaller, bigger or fine-tuned one, for example one tuned for their own city or campus.
• Free to run and fork: a student can host this for free and adapt it to their own surroundings, which is hard to do with a closed API.

Open models made it possible to build this without any budget, and that is exactly the kind of project that gets people outside.