Case study · 3 April 2026
Line QC: fewer claims, same takt time
Vision QC on a weld line. The model catches defects a shift cannot see in 0.8 seconds — without slowing takt.
The line welded faster than people could look. Claims came back weeks later, with a photo of a defect nobody on shift remembered. Adding inspectors broke takt. Adding cameras without a model produced footage, not decisions.
The problem
Surface defects and weld discontinuities escaped when the shift got tired. Sampling 10% did not protect the customer. Full manual inspection protected the customer and killed throughput.
What we deployed
A vision station at the end of the cell: cameras, lighting, a defect classifier, a semaphore for the operator. Rejects go to a queue, not a bin — the process engineer sees 20 frames, not 20,000.
The ERP gets a defect code and a batch tag. We did not build an “AI factory”. We wired inspection into the existing MES.
The result
Claims from this line dropped 41% in the first quarter after go-live. Takt held. False positives were high at the start — we walked them down weekly with the process engineer, not by “tuning in the cloud”.
If the model has no owner on the floor, it becomes a demo. Here the shift lead became the owner.
