CropScan

The problem

A farmer standing in a field with a sick plant usually doesn't have a good connection. A diagnosis tool that needs the cloud fails exactly where it's needed.

The app

• Point the camera, get a diagnosis in under a second, entirely on the phone.
• Disease recommendations ship with the app, so the full loop works with no network at all.

The training pipeline

• A dataset pipeline that validates file structure and labels, then shuffles and splits deterministically so runs are comparable.
• Transformations written as composable steps, which cut down on data bugs and training drift.
• Batching, prefetching and caching with explicit GPU/CPU selection to keep throughput up and memory predictable.
• Structured logs and per-epoch metrics, with checkpoints, plots and prediction samples saved for regression tracking.