I’m putting together a real-time inspection line that must catch both subtle textural defects and broader anomalies as products move past a USB industrial camera. The vision stack is Python-based, with YOLO as the starting architecture in PyTorch, and the finished model needs to run on an NVIDIA Jetson under tight latency constraints. Here’s the core flow I need help completing: continuous frame acquisition through the camera’s SDK (not a generic webcam wrapper), on-device preprocessing, YOLO inference accelerated with TensorRT, and immediate feedback—either through GPIO or a lightweight web dashboard—so operators can see the live image, bounding boxes, confidence scores, and flag any misses. Training scripts should let me fine-tune the model on newly collected samples, support anomaly detection add-ons, and export an engine that will drop straight onto Jetson with minimal fuss. I’ll provide initial data, target FPS, and acceptance thresholds; you bring the know-how to squeeze every millisecond out of the pipeline while keeping detection accuracy high. Deliverables: • Camera SDK integration module in Python • PyTorch training & export scripts, plus TensorRT-optimized engine • Real-time inference loop tested on Jetson (Linux) • Optional web UI for live display and operator feedback • Short setup guide so my team can reproduce, retrain, and deploy If you’ve already pushed YOLO or similar networks onto edge devices and dealt with industrial USB cameras, let’s talk—this project is ready to roll.