AI-Powered Crop Disease Lab

A production-shaped MLOps stack focused on plant-disease classification. PyTorch and TensorFlow models are trained through Kubeflow Pipelines, tracked and versioned in MLflow, and promoted from registry to a served endpoint. The lab covers the full lifecycle — reproducible training, model registry, deployment, and drift/accuracy monitoring — so new disease classes can be added without rebuilding the stack.

Crop-disease models are usually one-off notebooks with no path to reproducible retraining, versioning, or monitored deployment. As new diseases and regions appear, the model needs to be retrained and redeployed safely — which ad-hoc workflows can't support.

Training runs as Kubeflow pipelines that log metrics, params, and artifacts to MLflow; promoting a model is a registry tag change that syncs the served endpoint. Both PyTorch and TensorFlow backbones are supported. Deployed inference is monitored for accuracy drift, closing the loop back to retraining.

Planned — targets: reproducible training runs, one-command model promotion, monitored inference, and the ability to add disease classes without rebuilding the pipeline.

1. Phase 1

   Dataset & tracking

   Version the disease dataset and wire experiment tracking and a model registry in MLflow.

2. Phase 2

   Training orchestration

   Build Kubeflow pipelines for reproducible PyTorch/TensorFlow training.

3. Phase 3

   Deployment

   Promote registry models to a served inference endpoint.

4. Phase 4

   Monitoring

   Add accuracy/drift monitoring that triggers retraining.