A SMALL REASON TO STEP OUTSIDE
Look closer.
Then look up.
Give a photo to a tiny local learning model.
Take one color back into the real world.
JPG, PNG, WebP or AVIF · up to 12 MB · no upload to a server
Your scene, distilled.
Choose a color to notice on your next five-minute pause outside.
Find a second place this color lives.
Put the screen away.
Five minutes is a suggestion. No points. No streaks. No notifications.
03 / KEEP ONE DETAIL
What did you notice?
A sentence is enough. Observation notes are saved on this device only, when you choose.
What is the AI actually doing?
An open-source implementation of k-means++ unsupervised learning learns up to five CIELAB color clusters from your photo, on your device. Three deterministic restarts compete on training error. Every fifth sampled pixel is held out from fitting.
The error is color reconstruction distance (ΔE76), not confidence or ecological accuracy. Nearby pixels are correlated: this holdout is a diagnostic, not independent field validation. The model does not identify plants, species, objects, safe routes or edible things. Mission wording is a fixed template; the learned colors personalize it.
The demo is a generated illustration, not a claimed field test. Photos are downsampled and kept in memory, not added to notes. No GPS, accounts, inference servers or analytics. The initial web page is served online; download the source bundle to use the app offline.
Download the complete offline app + source (Apache-2.0) · Read the method and limitations