I need a hands-on computer-vision engineer to turn raw sports footage into structured insight. The system you build will ingest live or recorded matches, lock onto the ball and key players, keep them centered with an intelligent zoom, and raise flags the moment a goal is scored or a significant positional change occurs. I expect you to stitch together proven object-detection or tracking APIs with your own PyTorch, TensorFlow, or OpenCV code so we hit production-grade accuracy without reinventing wheels. The pipeline should: • track players, the ball, and other relevant objects frame-by-frame, • adjust the crop dynamically so the action stays in focus, • recognise high-value events such as goals or point scoring, as well as nuanced player movements and formations, • expose clean, well-documented endpoints or callbacks so downstream automation can act immediately on what the model sees, • run fast enough for practical use during real games. You’ll deliver commented source code, a reproducible training / inference setup, and a short demo video proving that the system can watch a full match highlight reel, detect at least three goal events, and output labelled clips or JSON events in real time. Optimize on a mid-range GPU so adoption is easy. If you have prior examples of sports analytics, dynamic zoom, or multi-object tracking at speed, include a link; seeing your work in action will fast-track the collaboration.