Caught 87% of line-defects before they hit packaging.

Two shifts a day of human QC inspectors still missed enough defects to trigger six-figure customer returns each quarter. We deployed a vision-and-vibration agent stack on the line and rebuilt the escalation playbook around it.

What was actually broken.

The plant ran two QC inspection stations per line, three lines, two shifts. Customer returns from the top three accounts averaged $1.7M a quarter, and the root cause was almost always a defect that left the line undetected. The plant manager had bought a vision system in 2021 that produced too many false positives — operators turned it off after six weeks. Inspectors were burning out and turnover on the QC floor was 38%.

What we did, and what we deliberately did not do.

We did not replace the inspectors. We replaced the bad vision system. We re-trained the cameras on the plant's own defect library, layered a vibration-sensor agent on the press itself (defects often originated upstream), and built a thresholded escalation: low-confidence flags went to the human inspector with the camera frame and the prior 30 seconds of telemetry overlaid; high-confidence flags stopped the line automatically. We deliberately tuned for low false-positive rates in the first 60 days so operators would trust the alerts. Trust took three weeks to rebuild and never broke afterward.

“The previous system cried wolf so often the operators muted it. This one cries wolf maybe twice a shift and they all check it. That is the only metric that mattered.”

What the numbers did, twelve months in.