Background
Experience
Engineered an end-to-end YOLOv8 object detection pipeline to automate the identification of competitor fixed broadband infrastructure from street-level imagery, replacing costly manual field surveys.
Improved overall model accuracy from a 0.522 to 0.763 mAP50 (+24%) through a data-centric approach: stratified dataset splitting, physics-aware augmentation, and high-resolution training at 1248px.
Resurrected failing minority classes from near-zero (~0%) to ~50% mAP50 via targeted synthetic data injection with a strict zero-contamination validation policy.
Built a geo-spatial inference engine integrating EXIF GPS extraction and offline reverse geocoding via BPS shapefiles, transforming raw detections into regional infrastructure intelligence reports.
