See it work, and see it decline.
Four demos, in your browser. Every number is computed from artifacts in the repository — including the ones that do not flatter us.
The chain working · held-out field photographs
A grape leaf with black rot, named correctly, from a field photograph.
Held-out field photographs of crops the app carries, with the answer the corpus recorded. How often the answers it puts on screen are right — including where that number is low.
The discipline · one photograph
It refused to diagnose, and said why.
One phone photograph. Trained disease models answered it confidently and disagreed with each other. The app showed none of them.
The data engine · measured, not asserted
A handful of your labels is worth something. One is not.
When the app refuses, it asks what you see. This measures what your answer is worth, on a crop it has no model for — against a colour histogram, a generic network and shuffled labels.
The frontier · turnable in your browser
A photograph flattens a plant. This one is metric.
269,833 gaussians from a phone capture, rasterised live in your browser. Leaf area, height and leaf area index become geometry instead of appearance — a flat photograph under-reports by 5.7×.
This demo renders a 269,833-splat 3D model in WebGL2 and is about 18 MB. It loads only when you ask for it.
Image credits
- Coffee. BRACOL — a Brazilian arabica coffee leaf dataset, CC BY 4.0.
- Pear. DiaMOS Plant — a field dataset for plant disease diagnosis, CC BY 4.0.
- Cotton. SAR-CLD-2024 — cotton leaf disease dataset, CC BY.
- Corn. Wiesner-Hanks et al. — maize northern leaf blight, expert-annotated, CC BY 4.0.
- Wheat. Zhou et al. — OFHB wheat Fusarium head blight, MIT / CC BY 4.0.
Our own photographs, screenshots and 3D capture. Dataset images are reproduced under the licences above, and each page repeats the licence beside the image it shows.
What these demos are, and are not
Research previews, labelled as such.
None of this is a field-validated diagnosis, and no page claims one. Disease inference ships fail-closed — it abstains until promotion evidence exists.