I wanted an excuse to look at plants instead of my phone. Wander is a plant guide where the phone's job is small: take one photo, get one answer, ask one follow-up, and put the phone away.
A Gemma model running on my own laptop names the plant and explains what it saw. Then I decide whether the answer is good enough to keep.
It started with a mistake
My first test photo was a sunflower, and Gemma got it right. My second was a pink flower with six spotted petals, and Gemma told me, confidently, that it was a Fuchsia. It was a lily.
That shaped the app. A small local model is useful, but it can be wrong in a convincing voice. So Wander keeps the human in charge:
- Nothing counts until you confirm it, and you can edit the name first.
- Every answer shows its doubt: a confidence score, the features it used, and look-alikes it may be confusing.
- You can cross-examine it. The Ask tab sends the same photo back and lets you challenge the answer, for example "Why not a lily?"
- The model is told never to call a plant safe, and every result carries a warning not to eat or handle plants based on an AI guess.
What it does
- Find: photograph a plant and get a name, confidence, a short summary and visible features.
- Map: each confirmed find appears where you took the photo, or where you drop a pin.
- Collection: your species, catalogued like a herbarium, with every sighting.
- Badges: milestones like first find, five species and an early-morning sighting. These are plain rules in code, not AI.
Why open-weight mattered
Identification and follow-up questions run through Ollama on my own laptop (RTX 3050, 16 GB RAM) using gemma4:e2b-it-qat, a 4.3 GB download. The photos are processed on my machine, there's no cost per photo, and the identification prompt, safety rules and chat behavior are all text I can change in my own code.
It isn't fully offline. The photos and the AI model stay on my laptop, but the map background loads from OpenStreetMap, and sighting records are stored in MongoDB, which has to be reachable when the app runs.
How it's built
React and Tailwind on the front, an Express server for uploads, MongoDB and badges, and a small FastAPI service that makes Gemma answer in strict JSON. My phone connects to the laptop over the local network. Most of the code was written with AI coding assistants. I directed the design, ran every service and tested it on photos of plants.
What's next
- A proper walk. So far I've tested Wander on photos of plants on my laptop. The next step is taking it outside and counting how often it's right.
- Offline map tiles. The map background needs internet today, so I want to cache map tiles for the area I'll walk in.
- A specialist plant model for naming. Gemma sometimes names a flower wrongly with high confidence. I'd like to test a dedicated plant classifier for the naming step and keep Gemma for the explanations and questions.
GitHub: https://github.com/AMAN240310/Wander
Prize category: Best Use of Gemma
Top comments (0)