I built a fully offline nature field guide with Gemma 3

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

What I Built

Trailside is an on-device nature field guide. Point your camera at a plant, bird, insect, or fungus and get field notes in seconds — no signal, no cloud, no account. Everything runs locally with a small vision model, so it works on the trail, in the woods, or anywhere the cell bars disappear.
It is built for people who want to look closer at the living world instead of scrolling past it: hikers, gardeners, kids on weekend walks, anyone who has ever wondered “what is that?” and wished for a quiet answer in their pocket. You photograph something, optionally add a quick note of what you already know, and Trailside returns structured field notes. Save the ones that matter to a personal LogBook so the memory stays with the place, not the feed.
The goal is simple: get people off the screen and back into the world — with just enough help to notice more of what is already around them.

Demo

Vidoe Demo

Code

https://github.com/Gavinduachintha/Trailside

How I Built It

Trailside is built around Ollama running the open-weight Gemma 3 vision model (gemma3:latest) entirely on-device. There is no cloud API and no remote inference — the model lives on the user’s machine, so identification works offline on the trail.
The app is a Streamlit front end that handles camera capture or file upload, lightly processes the image with Pillow (EXIF orientation, resize, JPEG encode), then streams the photo and an optional user hint into Ollama’s chat API. A system prompt and a small prompt builder shape the model’s reply into readable field notes. Responses are streamed token-by-token into a journal-style panel so the experience feels immediate.
Saved sightings (image + notes + optional hint) are stored in a local database via a thin config.db layer, and the LogBook view lets users browse, open, and delete entries. The whole stack — UI, local model, and storage — stays on the device so the tool remains useful where signal is weak or absent.

Why Does Open Innovation Matter?

Open innovation is what makes Trailside possible as a real field tool instead of another phone feature that dies when the bars disappear.
Because the model is open-weight and runs locally through Ollama, every photo stays on the device. There is no upload, no account, no usage quota, and no dependency on a remote API that could change price, rate limits, or terms overnight. That privacy and reliability matter when you are identifying something in the woods — not because the data is sensitive in a corporate sense, but because the tool should work where you actually need it: offline, on a trail, with whatever hardware you already have.
A closed vision API would have forced a constant network connection, sent images off-device, and tied the project to someone else’s pricing and availability. Open models and local inference remove that gate. They let a small, single-purpose app treat the model as a quiet companion on the path

My Agent Session

Prize Categories

Best Use of Gemma — Trailside is built around Gemma 3 running locally through Ollama. The model does the full multimodal identification: it receives the photo (and optional user hint), streams the field notes, and never leaves the device. No other AI or partner runtime is involved.