PatchWatch: I Built a Voice-First AI That Turns a 10-Minute Walk Into a Biodiversity Survey
The problem
Going outside doesn't have to mean hiking for hours or travelling to a national park. A campus lawn, a roadside tree or a neighbourhood garden can become an opportunity to notice the natural world.
But if you don't know where to start, even a short outdoor break can feel aimless. I wanted to build something that makes those small experiences more purposeful without keeping people glued to their phones.
What I built
PatchWatch is a voice-first biodiversity observation tool. Users choose a habitat and the time they have available, receive three short observation missions, and then head outdoors.
They can hear the instructions, put their device away, record what they noticed after the walk and save a structured field report. The records can be exported as CSV, making it possible to compare observations across different walks.
Why open-weight AI matters
PatchWatch uses Gemma through Ollama for mission generation and field-note summaries.
Running the model locally means that the core AI workflow doesn't need to send a user's field notes or precise location to a hosted language model. After downloading the model, local inference can run without an internet connection while Ollama is available.
ElevenLabs is an optional voice layer that turns instructions into an audio guide. The generated MP3 can be saved ahead of a walk for offline playback. Unlike the local Gemma workflow, generating new speech through ElevenLabs requires a network connection.
I wanted the core AI to remain local and replaceable instead of making the entire experience dependent on one hosted AI endpoint.
How I built it
- React and Vite for the interface.
- Node.js and Express for the backend.
- Gemma 3 1B and Ollama for local AI inference.
- ElevenLabs for audio generation.
- localStorage for field notes and walk history.
- CSV export for taking observations beyond the application.
My real-world test
I tested PatchWatch during a [duration]-minute walk in [general type of location].
I completed [number] of the three missions. One observation that surprised me was [actual observation]. The main limitation I noticed was [actual limitation], which I would address next by [planned improvement].
What I learned
An outdoor AI tool should make the screen the shortest part of the experience. The point isn't to turn a walk into another session of endless app interaction. It's to offer a little guidance, let the user explore and create something useful from what they noticed.
My next step is to make repeat observations easier to compare and let campus communities build their own anonymized biodiversity records.
Prize categories entered: Best Use of Gemma; Best Use of ElevenLabs; Overall Challenge — only where the completed implementation satisfies each category's requirements.
Top comments (0)