Vision and sensor models on the line that catch defects and predict failures before downtime. Models are proven on a test line, but the hard part is production: inference at the edge, continuous sensor data, and retraining as lines and products change.