• Developed an HRNet-based keypoint detection model for pitch calibration and homography estimation, enabling transformation of image coordinates into real-world space and computation of player distance, speed, and possession metrics. • Trained video action recognition models for sports sequences, achieving 82% accuracy in tactical event classification. • Implemented a human-in-the-loop pipeline, reducing detection and tracking errors by 80%. • Deployed real-time AI models using ONNX Runtime, FastAPI, and Docker for scalable and low-latency inference, reducing match analysis time by 80%.