Jani 1
Lina kutu. Hili ni jani moja. Kabla ya kutumia pesa, onyesha jani hili kwa afisa ugani au ushirika wako.
Je, jani hili la kahawa lina kutu?
Huu ni mfano tu. Hizi ni picha nne za mfano, na jibu la kompyuta kwa kila moja.
Jani 2
Halina kutu. Hili ni jani moja tu. Kagua majani mengine. Ukiwa na wasiwasi, onyesha mtu.
Jani 3
Sina uhakika. Hii haimaanishi kuwa jani lina afya. Onyesha jani hili kwa afisa ugani au ushirika wako.
Jani 4
Sina uhakika. Hii haimaanishi kuwa jani lina afya. Onyesha jani hili kwa afisa ugani au ushirika wako.
Ni msaada tu. Mtu ndiye anayeamua.
For reviewers: how it works and its limits
- What it does. One photo of one coffee leaf goes in. One of three fixed answers comes out: rust, no rust or not sure. It never writes free text, so every answer can be checked. A person makes the final call.
- Model. Gemma 4 E2B, fine-tuned with LoRA on 128 real BRACOL photos.
- Evidence. On 300 held-out Uganda farm photos, half with rust, it answers 72% of them (about 7 in 10) and is right on 96.5% of those (about 19 in 20). For the other 28% it says “not sure”.
- This web demo is a presentation. The camera is simulated: it takes no real photo and runs no model. The four photos and their answers are saved from the real model, and the short “looking at the leaf” step only shows how the real flow feels. Every simulated screen says “Example only”. Nothing is uploaded, and the page never asks for the camera, microphone or location.
- The real tool is meant to run on a phone with no internet. A 5.2 GB phone file exists. It has not been tested with farmers, and at that size it must be side-loaded.
- Offline. After one visit with internet, this page works with no connection. A plain-HTML copy of the four answers shows when scripts do not run (for example on Opera Mini). On iPhone, add the page to the home screen, because Safari can delete saved site data after seven days.
- Language. Swahili, in text and voice. The voice is machine-made (Meta MMS-TTS, CC-BY-NC 4.0) and a native speaker has not checked the Swahili. Because the answers come from a fixed list, a new language needs a few short phrases and recordings, not a new model.
- Accessibility. Designed from a desk study of low-literacy and low-connectivity users (see Design study). Automated checks pass. They do not replace tests with farmers, which have not happened yet.
- What the data does not cover. BRACOL is Brazilian Arabica leaves on clean backgrounds. Our field test is Uganda only, with small close-ups and no severity labels. We have not tested other regions, other crops, or blurred and no-leaf photos. One leaf does not show the state of a whole plant.
- Version.