Computer Vision
Sentinel
Real-time defect detection running at 60 FPS on a Jetson Orin Nano — quantised, distilled and deployed to 140 production lines.
Constraints first
The camera was fixed, the compute was a 15 W board, and a missed defect cost more than a false alarm. That ordering decided every architectural choice.
Getting to 60 FPS
Knowledge distillation from a ConvNeXt teacher into a small custom backbone, then INT8 post-training quantisation with a 2,000-image calibration set. Accuracy loss from quantisation: 0.4 points of recall.
Drift is the real problem
Models degrade because the world changes — new supplier, different lighting, a cleaned lens. A weekly embedding-distribution check against a reference set catches drift long before the accuracy metric does.
Next