How we validate every AI diagnosis

We receive a photo. We return the name of a pest, an approved product and a dose. In between there is a chain we deliberately do not fully automate.

The AI proposes a hypothesis, not a verdict

The model receives the photo and the declared crop type. It returns the three most likely identifications, each with a confidence score. It never returns a single answer: a yellowing leaf is compatible with a deficiency, waterlogging and three diseases.

When the top hypothesis falls below a confidence threshold, or when the first two are too close together, the ticket is flagged as priority in the agronomist’s queue.

The agronomist decides, and can ask again

The agronomist sees the photo, the model’s hypotheses, the crop, the municipality and the last seven days of weather. That last part matters: a brown spot after three humid nights does not read the same as the same spot after ten days of dry wind.

They confirm, correct, or ask for a second photo — most often of the underside of the leaf, which on its own settles half the ambiguous cases.

The prescription is filtered, not generated

Once the pest is confirmed, the treatment does not come out of the model. It comes from the catalogue of approved products, filtered on three hard criteria: authorisation on the crop concerned, efficacy against the identified pest, and a pre-harvest interval compatible with the plot’s stage.

If no product passes all three filters, the app proposes nothing and the agronomist writes a manual reply. We would rather give a slow answer than an illegal recommendation.

What we measure

Every agronomist correction is recorded as a gap between the model’s hypothesis and the retained diagnosis. Those gaps feed retraining, and more importantly they get read: a pest that is consistently misidentified points to a hole in the dataset, not to a bad day for the model.

The rule that holds the whole thing together has not changed since day one: no advice leaves without an agronomist having signed it.