Amaizing

Maize disease is diagnosed late, by which time the response is salvage rather than treatment. The diagnosis itself is a visual judgement an expert makes in seconds, and the expert is rarely in the field.

Getting that judgement onto the phone already in the farmer's hand is worth more than improving its accuracy past a point, which shaped every decision about how it was deployed.

The training pipeline, the preprocessing stages, the experiments and the saved models, and an explainability layer that produces a heat map showing which part of the leaf drove the classification. A prediction without that is an assertion, and an agronomist has no reason to accept it.

An API with readiness and health probes that account for a model that loads lazily, an installable progressive web application in front of it, and a containerised deployment behind a reverse proxy with certificates handled automatically.

A convolutional network fine tuned to distinguish healthy leaves from the three diseases that matter, over a preprocessing pipeline of segmentation, normalisation, and augmentation that is applied identically at training and at inference. That identity is not a detail: a model that sees differently shaped inputs in production has quietly stopped being the model that was validated.

The deployment sidesteps inference hosting entirely. The trained weights live in ordinary file storage, the service downloads them once on its first readiness check into a persistent volume, and every prediction after that runs locally. There is no per prediction cost and no external dependency in the request path.

Lead developer. The model, the preprocessing and training pipeline, the explainability layer, the API, the web application, and the deployment.

Deployed and running in production. It is one of the real services running through SlideOps on its own infrastructure, which is how that platform's deploy and reverse proxy path came to be proven outside of a test.