Tim Hortons Cup Object Detection Pipeline
Published:
Built and deployed an end-to-end classification/detection pipeline, serving model inference through Flask and FastAPI endpoints, with MLflow, DVC, Prometheus, and Grafana for tracking and monitoring. Conducted precision/recall-oriented failure analysis using CAM overlays, confusion matrices, scenario heatmaps, and UMAP, improving F1 from 0.942 to 0.959. Deployed a final model achieving 95.4% image-level accuracy and F1 = 0.967, monitored with real-time drift detection.
Tools: YOLOv8, Flask, FastAPI, MLflow, Prometheus, Grafana, DVC
