Production Computer Vision

Learn from the missed cases

I rebuilt a production detection workflow so we could improve the model and inspect where it failed. The work combined model development with a more useful labeling and evaluation process.

I examined missed detections, improved the training workflow, and compared model behavior against labeled examples. That led to better recall and a clearer basis for deciding what to work on next.

The most useful part was the feedback loop between errors, labels, and training. A model change mattered when it addressed the cases people were actually encountering.