Things I've built

Production-grade MLOps pipeline for real-time African wildlife detection. GPU-accelerated Faster R-CNN served via TensorFlow Serving and gRPC, Flask frontend, full Prometheus + Grafana observability stack, containerised with Docker Compose. Custom Prometheus metrics track inference time, detection counts, and confidence scores with production alert design for data drift and model degradation.

Rigorous comparative study of two-stage vs one-stage object detection architectures on African wildlife camera trap data. YOLOv8-L achieved 90%+ mAP@0.50:0.95 at 82.44 FPS — 11.5x faster than Faster R-CNN ResNet-101. Faster R-CNN achieved AP@0.50 of 95.5%, making it superior for scientific census work. Stratified data splits and dual isolated environments for scientific rigour.

Multi-class ML classification pipeline mapping avian skeletal measurements to six ecological niches. EDA on 420 specimens with feature correlation analysis and missing value handling. Random Forest selected as best model — 74.70% test accuracy, Macro F1 of 0.67 — evaluated against SVM and Logistic Regression via AUC, ROC curves, and confusion matrices.