The finish defect that no longer reaches the customer

The problem

At line speed, a scratched cabinet looks like a good one.

100% human visual control is impossible at real line speed. Sample-based inspection leaves gaps that end up as warranty claims or distributor returns.

• Checking 100% of cabinets by eye is impossible at real line speed. Inspection is done by sampling, and the sample is only as good as the inspector's last hour. At the end of a ten-hour shift, that hour is not the best one.
• Scratches, dents, orange peel, runs and thin paint pass in the gaps between samples, especially at the end of a long shift or after a color change. A cabinet missed there will reach a customer's kitchen with the defect in plain sight.
• Those units come back as finish-defect claims from distributors or as warranty returns, with freight both ways and a replacement unit on top. A dented cabinet that reaches the customer's kitchen also costs a service visit and a bad review.
• And every returned cabinet costs far more than the scrap or rework it would have been on the line, before packing.

How it fits the IRIS system

Edge — a local CNN over the line, inference in milliseconds.

Edge camera over the line + local CNN trained on good/bad examples of the specific cabinet. Millisecond inference, no cloud dependency.

The model is trained on good and bad examples of your own cabinets, under your line's lighting. It does not depend on the cloud or the plant network, so it keeps up with the conveyor and never holds it. The inspector stops sampling and starts deciding the borderline cases. That is where their experience is worth most. Their decisions on those cases are what keep improving the model.

See the full IRIS architecture →

Before and after

Sampled finish checks versus 100% Edge inspection

Aspect	Today	With iLEAN Edge
Cabinets inspected	A sample	Every unit
Where the defect is caught	At the distributor	Before packing
Inspector fatigue in white goods	Grows over the shift	Irrelevant to the camera
Decision time per cabinet	—	Milliseconds, local
Defect history	Anecdotes	Images by shift and model
Borderline finish	Passed or failed by feel	Escalated to quality

Partial sampling → 100% unit control. Finish-defect claims → sharp reduction.

Impact estimate

Impact estimate — to be validated with your numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 5-12 months, depending on your current rejection rate.
• From partial sampling to 100% unit control at real line speed. Every cabinet, every shift, every color. The inspector's role moves from sampling to judging the hard cases.
• A sharp reduction in finish-defect claims and returns from distributors, with the freight and replacement units they bring.
• And a defect history by shift and model that points back at the paint line or the handling step causing it. Fixing the cause upstream is worth more than catching the defect downstream.

Estimated payback 5-12 months depending on current rejection rate. *Estimate to validate*.

And the fair question from the production manager

"What if it rejects good cabinets?" — false positives are the real risk of any vision system, so the model is trained on your own cabinets and lighting, not a generic one. Classifying a known defect on a known surface is an anchored task, where the best models drop below 1.5% error [1], and borderline units go to a quality inspector rather than straight to scrap. Each decision the inspector makes feeds back into training. The camera does not replace the inspector's judgment; it replaces the sampling that today decides which cabinets nobody looks at.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.