Application for defect recognition on production line via camera

We have a small manufacturing operation and quality control is currently completely manual: an operator takes each product, turns it under a lamp and decides whether it's good or defective. People get tired, lighting changes throughout the shift, and toward the end of the day obvious defects start slipping through. We need a computer vision-based application that looks at the line through a camera and marks defects itself so a person only confirms disputed cases rather than manually checking every single unit from scratch.

The task is to actually apply a trained CV model to live quality control, not just label a dataset. The camera should capture each product as it passes, the model should classify it as good or defective, and if possible identify the type of defect: surface scratches, missing or incorrectly installed parts, print errors — smudged, offset or defective labels. Each rejected item needs to be photographed, recorded with a timestamp and defect category and shown on the dashboard so the shift supervisor sees the count of good and defective items, the defect rate over time and can open the saved image of any marked item and double-check the decision.

We will provide samples of photos of good and defective products for training and tuning, and we're ready to discuss camera choice, lighting and framework by your recommendation. For us it's important to have a working pipeline that will actually run in the shop: reliable frame capture, reasonable thresholds so we don't get overwhelmed with false positives, and an understandable log that can be reviewed. Please describe how you'll collect and label training images, what model architecture you propose, how you'll measure precision and recall on our parts and how the finished application will be installed and maintained by non-technical staff.