obaidullah-faruk/african-wildlife-mlops

An MLOps project for detecting buffalo, elephant, rhino, and zebra in images. # Wildlife MLOps Learn one object-detection lifecycle. The model detects buffalo, elephant, rhino, and zebra. ## Architecture ```mermaid flowchart LR data[Versioned wildlife dataset] --> validate[Dataset validation] validate --> train[Baseline training] train --> validation_metrics[Validation metrics] train --> checkpoint[best.pt checkpoint] train --> mlflow[Local MLflow] validation_metrics --> candidate[Release candidate] checkpoint --> candidate candidate --> package[Package: checkpoint + inference code] candidate --> approval[Human approval] approval --> sealed[One sealed test evaluation] sealed --> service[FastAPI service\nloads one pinned candidate] client[Local prediction client] --> service service --> response[Prediction JSON\nversion + SHA-256 checksum] response --> rollback[Manual A → B → A rollback evidence] service --> service_metrics["/metrics/ endpoint"] service_metrics --> prometheus[Prometheus\nscrapes every 5 seconds] prometheus --> grafana[Grafana\nrequest, latency, error, and distribution graphs] response --> samples[Sampled prediction metadata\nno raw images] labels[Later ground-truth labels] --> quality[Sampled precision and recall report] samples --> quality failed[Failed candidate start] --> recovery[Recovery health evidence] service --> recovery ``` ## Setup ```sh make bootstrap make doctor ``` ## Learn the data ```sh make data-download make data-validate make data-visualize ``` Open `artifacts/data-preview/train.png`. Check that the boxes match the animals. ## Learn training ```sh make predict-pretrained make train-overfit make train-smoke make train-baseline ``` Each training command creates a new directory under `artifacts/`. Open its `results.csv`, `run.json`, and `weights/best.pt`. `train-overfit` uses a few images twice. Inspect whether its loss falls. `train-smoke` proves the full training path works quickly. `train-baseline` uses the whole training split. ## Track a run Copy the example credentials once, then start local …