About
Explore more. Scroll less.
Most apps want more of your screen time. WildQuest uses the screen to get you outside, then asks you to put it away.
How it works
- Pick an adventure. Three, five or seven small missions, like finding a leaf or photographing the sky.
- Photograph what you find. Use your camera or choose a photo.
- AI compares it on your device. The photo is scored against short text descriptions: your target, look-alikes, and unrelated things.
- A clear win counts. If your target leads by enough, the mission is verified. If not, you can retake the photo or log it yourself.
The open-weight model
WildQuest runs CLIP ViT-B/32, an open-weight vision-language model from OpenAI (MIT license), in the ONNX build Xenova/clip-vit-base-patch32, using Transformers.js and ONNX Runtime Web. It runs in a background worker on your device.
CLIP gives relative similarity scores, not the chance that you’re right. It can be fooled by look-alikes, bad light or blur. That’s why unclear results are called “not sure”, and why manual logging exists and is labeled as manual.
Why open weights matter here
- Privacy: the model comes to your photo, so your photo doesn’t go to a server.
- Parks have bad signal: weights can be saved on the device.
- No per-photo cost: it runs on your own hardware.
- Inspectable: the model, runtime and mission prompts are all open.
Credits
CLIP (OpenAI, MIT) · Xenova ONNX conversion · Transformers.js (Apache-2.0) · ONNX Runtime Web (MIT) · Next.js, React, Tailwind CSS, Dexie, Lucide (MIT/ISC) · Inter (SIL OFL). Built for the DEV Hacktoberfest “Touch Grass” challenge.