Side Quest: An AI That Wants You to Put Your Phone Down

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

What I Built

I know I should go outside more. What usually stops me isn't time, it's not having a reason. I'd open my laptop 'for five minutes' and look up an hour later.

Side Quest gives you that reason. You tell it how long you have and where you are, and a local AI gives you one small quest:

• three things to find (like "a bird you can hear but not see")
• a two-minute sit spot (stop and listen)
• a time to turn back

Then the screen goes dark and says "Phone away." When you're back, you snap your finds, the AI checks your photos, and it writes a short journal entry about your walk.

The only number it shows you at the end is minutes outside vs minutes on screen.

Demo

Live demo: side-quest-vpar.onrender.com

The live demo runs on Render's free tier, which doesn't have enough memory for Gemma, so it serves quests from the built-in offline deck. The full experience, with Gemma writing quests and checking photos, runs locally and is shown in the screenshots below. If the demo takes a minute to load, the free server is waking up.

Code

https://github.com/ananya-in-tech/week-1-touch-grass

How I Built It

• Model: Gemma 4 (E2B), running locally through Ollama. No cloud API.
• Backend: Python + FastAPI, SQLite for storage.
• Frontend: a small mobile-friendly web page with no build step.

Gemma does three jobs:

1. Writes the quest. It returns structured JSON that the app validates. A small rules layer keeps quests safe (no eating, touching animals, climbing or going near water) and makes sure there's always a real time limit.
2. Checks your photos. Gemma's vision looks at each photo and says whether it matches the find. Before that, the app removes all location data from the photo.
3. Writes your journal. It uses only what you actually found and noticed, so it doesn't make things up.

If the AI is off, Side Quest still works using a built-in set of quests.

Why Open Innovation Matters Here

• Privacy: photos from a walk reveal where you live and where you go. With a local model they never leave my computer, and the location data is stripped before the model even sees them.
• Free to use: no API key and no per-photo cost, so there's no reason to ration walks.
• I can shape it: with an open model I could tune it for where I live. For example, my first quests said "autumn leaves", but October in India is post-monsoon, so I made the seasons region-aware.

What Went Wrong

• Early quests used the wrong season for India. I fixed it with region-aware seasons.
• The AI sometimes gave vague turn-back times like "when the shadows lengthen". The app now always sets a real time limit.
• On a laptop with no GPU, the first photo check took about 2.5 minutes while the model loaded. The app now loads the model when it starts.

Taking It Outside

[Describe your real walk: where, how long, what you found, what Gemma said about your photos, and your outside-vs-screen numbers.]

Prize Category

Best Use of Gemma: Gemma 4 runs locally and powers quest writing, photo checking (vision) and the journal.