Coy

The wild side of campus. Now in focus.


Inspiration

UC Santa Cruz isn't just a campus. It's a forest.

Deer at dusk. Jays in the oaks. Banana slugs on the trail after the first rain.

Thousands of students walk past it every day. But they don’t stop to admire this wonderful fauna.

We thought that should change. Because every one of those moments is more than a moment. It's a data point. A species, a place, a time. That's how scientists track how nature moves and changes. Community science has shown that anyone with a phone can contribute. All it takes is making it effortless.

And making it fun.

What it does

Meet Coy.

Point. Shoot. Know.

Take a photo of something wild, and Coy tells you what it is. In under a second. It knows 34 species that call campus home. Mammals. Birds. Native trees. Insects. And yes, the banana slug.

You get its best three guesses, each with a confidence you can trust. Show it something it doesn't know, and Coy does something remarkable. It tells you it isn't sure.

Name your find. Pin it on the map. Share it.

Every sighting joins a living campus feed. Rate it. Comment on it. Watch the map fill with life. The Species tab becomes a field guide written by everyone. The Ranked tab celebrates the best finds.

It's a nature walk. It's a scavenger hunt. It's a research tool.

It's all three.

How we built it

33,358 photos. Every one earned its place.

Each image comes from iNaturalist, verified by the community and openly licensed by its photographer. We removed every duplicate and near duplicate. We credited every photographer.

Built on a model that has seen life itself.

We started with BioCLIP 2, trained on more than 200 million images across the tree of life. Then we taught it our campus. Using Google Colab, we trained it on our 34 species and fine-tuned its final layers.

Tested the hard way.

We tested only on photographers the model had never seen. No shortcuts. No memorized backgrounds.

95.5% accurate. 98.3% in the top three.

And calibrated, so when Coy says 90%, it means 90%.

Fast. Light. Private.

Coy identifies a photo in about 80 milliseconds. It ships as a 99 MB adapter, not a multi-gigabyte model. Every uploaded photo is stripped of hidden location data, so the only place stored is the one you choose to share.

A Python and FastAPI backend. A native SwiftUI app. Google Maps for every sighting.

Challenges we ran into

Nature doesn't pose.

Bobcats, foxes, and raccoons come out at night. Most of their photos are grainy trail camera shots. Two kinds of gray squirrel look nearly identical. These were our hardest species. We report them openly.

Our mascot was camera shy.

There just aren't many openly licensed banana slug photos. We trained on 358 instead of 1,000. It still scored perfectly on its test set.

Teaching a model to be honest.

A model trained on 34 species always wants to answer with one of them. Even for a flower. So we calibrated it, and designed Coy to say "I'm not sure." Because a confident wrong answer is worse than no answer at all.

Accomplishments that we're proud of

A model that recognizes what it has never seen before. And admits what it doesn't know.

From photo to mapped sighting in a few taps.

Every image openly licensed. Every photographer credited.

What we learned

Accuracy is only half of trust. The other half is honesty.

The hard way to test a model is the right way. Splitting by photographer cost us a few points of headline accuracy. It gave us a number we can believe.

What's next for Coy

Works without signal. On-device identification with Core ML, for the deep redwoods.

Real science. Export every sighting to iNaturalist and GBIF, so a walk to class becomes a contribution to biodiversity research.

Protect what's rare. Automatically blur the locations of sensitive species.

More to discover. Seasonal maps to watch slug season roll in with the rain. More species from the Campus Natural Reserve. Challenges to find all 34.


The wild was always here.

Coy helps you see it.

Built with

Swift · SwiftUI · Python · PyTorch · OpenCLIP · BioCLIP 2 · FastAPI · SQLite · Google Colab · Google Maps SDK for iOS · Hugging Face · iNaturalist API

Built With

Share this project:

Updates

Submission history