I have continuous video footage from live inspection of steel flat bars strips our production line and need a complete deep-learning pipeline that flags surface, structural, and functional defects in real time. The raw videos are already labeled by timestamp; frame-level annotation may still be required for optimal accuracy. Scope • Design and train a deep neural network—CNN, transformer, or hybrid model—that detects all three defect categories directly from video streams. • Implement preprocessing (frame extraction, augmentation, ROI isolation) and post-processing (tracking, alert generation) in Python using libraries such as PyTorch/TensorFlow and OpenCV. • Optimise for inference on an on-premise GPU; latency under 200 ms per frame is the target. • Provide clear metrics: precision, recall, F1, and confusion matrices on a held-out validation set. • Package the final solution with a lightweight REST or gRPC endpoint so the in-house engineering team can call it from our SCADA system. Deliverables 1. Source code repository with clean, documented modules. 2. Trained model weights and a reproducible training script. 3. Step-by-step deployment guide (Docker-friendly). 4. Short report summarising data preparation, architecture choices, and evaluation results. Acceptance Criteria • ≥95 % F1 on the labelled test set for each defect type. • End-to-end demo on a sample video showing automatic detection and bounding-box overlay. • No proprietary licenses; solution must be fully redistributable within the company. Once these items are ready I can move straight to integration, so please keep the code modular and well commented.