I am building an AI-powered CCTV application that scans live in-store video and raises an on-screen alert the instant it spots behaviour linked to shoplifting. From the first release the model must recognise: • Unusual hand movements near products • Frequent glancing around • Lingering in an area for too long By tracking hand, head and eye direction the system should outperform human observers and keep false positives low. It will be deployed across supermarkets, boutique stores, electronics stores and must remain adaptable to any other retail layout without requiring new cameras. The workflow I have in mind: incoming feed → OpenCV (or equivalent) preprocessing → deep-learning model (YOLO, Detectron2, TensorFlow/PyTorch—whichever you favour) → real-time overlay that highlights the suspect and pops a visual alert in the guard’s monitoring software. No audible alarm is needed right now. Deliverables 1. Trained model and reproducible training pipeline 2. Low-latency inference service that runs on standard Windows or Linux hardware 3. Operator dashboard showing the flagged camera, bounding boxes and a timestamped event log 4. Deployment script or installer plus concise documentation 5. Short demonstration video from a test environment Acceptance criteria: sub-second detection on a 1080p stream, F1-score ≥ 0.80 for the behaviours listed and stable eight-hour continuous operation without crashes. If you can deliver a turnkey solution that meets these metrics, I’m ready to set milestones and move ahead.