Crop-health AI

The delivered flow let a farmer photograph a leaf offline and receive the diagnosis when the phone reconnected. That separated evidence collection from connection availability, rather than require both to happen at the same moment.

The product question extended beyond whether a model could identify a condition. A usable journey needed an obtainable photograph, a path back to the answer and a clear boundary around what that answer established.

Crop-specific field playbooks and an OpenAI API-powered conversational layer were intended to connect the diagnosis to practical guidance. The records do not establish their production release. They are not presented as a delivered assistant or a measured reduction in support work.

Low-end phones and inference efficiency were delivery constraints alongside connectivity. The account does not include a publishable retry/recovery trace or a defined low-confidence threshold, so those details remain proposed validation questions rather than historical controls.

The project included image curation, labelling and an agronomist review loop. Specialists helped establish the meaning of labels while engineering and modelling colleagues delivered the technical capability.

My product remit connected that work with the farmer’s help-seeking flow: what evidence could be collected, when an answer could return and how the result could be understood. It does not imply that I personally trained the models or created every annotation.

The reported dataset covered more than 500,000 images across fourteen crops. That demonstrates the described data scope, not class balance, held-out performance or suitability for every image quality and field condition. The source does not contain a publishable requirement example from the agronomy review.

For an AI product like this I would examine the point of action: can someone collect usable evidence, recover from a failed connection, understand the answer’s limits and take an appropriate next step? Those product decisions belong beside the model output, rather than follow it as interface polish.