The most expensive part to scrap is the one already finished

The problem

The most expensive part to scrap is the one already finished.

Fabric wrinkle, bubbling, incipient delamination, mold mark, stain, out-of-tolerance trim, badly seated speaker ring, missing sun visor or grab handle, unmated connector. With a hundred active variants, each with its own component combination, one hundred percent human visual control is not sustainable and sampling leaves gaps by definition. Besides, the headliner is the most expensive part to scrap in the whole plant: when it fails, it already carries the formed substrate, the lamination, the trimming and the full assembly. All the added value is inside what gets thrown away. And the one that slips through does not reach an end consumer: it reaches the customer's assembly line, where an operator sees it and stops the line.

• Wrinkle, bubbling, incipient delamination, mold mark, stain, out-of-tolerance trim, badly seated speaker ring, missing visor or handle, unmated connector.
• With a hundred active variants, each with its component combination, one hundred percent human visual control is not sustainable and sampling leaves gaps by definition.
• When it fails, the part already carries the formed substrate, the lamination, the trimming and the full assembly: all the added value is inside what gets thrown away.
• And the one that slips through does not reach a consumer: it reaches the customer's assembly line, where an operator sees it and stops the line.

How it fits the IRIS system

Edge — vision with local inference, at the three points where the defect appears.

Edge — CNN vision with local inference.

The image does not leave the plant: no cloud latency, no bandwidth cost, no customer product circulating outside the site. The suspect part leaves the sequence and a person decides its final destination.

• Cameras at the lamination outfeed, at the trimming center and at the final assembly station, with lighting designed to reveal surface defects backlit.
• Inference on a local plant GPU, in milliseconds per part.
• One model per defect family, trained with real good and bad parts from your own production, not with generic catalogs.
• The detection is cross-referenced with the active variant to also verify that every component that specific variant must carry is present.
• The suspect part leaves the sequence and is flagged for human review; the person decides its final destination.
• The image does not leave the plant: no cloud latency, no bandwidth cost, no customer product circulating outside the site.

See the full IRIS architecture →

Before and after

Sampling vs. one hundred percent control at three points

Aspect	Sampling control	With iLEAN Edge
Coverage	One or two percent	100% at three points in the flow
Where it is detected	Sometimes, at the customer	Before the rack
Control performance	Drops from the fourth hour	Constant all shift
Component presence	Depends on the variant and the eye	Verified against the active variant
Value of the scrapped part	Finished, with all the value inside	Detected at the first process
Where the image goes	There is no image	It stays in the plant

From sampling one or two percent to controlling one hundred percent of parts at three points in the flow. From a defect found at the customer to a defect found before the rack.

Impact estimate

Impact estimate for your plant — to validate against 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.

• Headliner: a large, light, matte part whose defects show up backlit and at an angle, not head-on.
• Pilot on one point of the flow, with real good and bad parts from your own production as training material.
• Indicative payback between 5 and 12 months depending on the current scrap and rework rate.
• What gets counted is the reduction in PPM to the customer and the risk of paid sorting, which is where the big cost sits.

Estimated payback 5 to 12 months depending on the current scrap and rework rate, with reduced PPM to the customer and reduced risk of paid sorting. *Estimate to validate against the scrap and claims history*.

And the fair question from the production manager

“What if the model pulls good parts out of the sequence?” — the false positive rate is measured from day one of the pilot and is a tunable parameter: it is calibrated with good and bad parts from your own production. A model trained on a generic catalog does tend to over-reject; one trained on your fabric, your lighting and your variants does not. And the suspect part is not scrapped on its own: it is flagged for human review.

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