The trim defect stays at home

The problem

At cadence the human eye tires, and the criteria drift.

At line cadence, with variants changing every few units, the human eye tires and lets things through. One hundred percent visual inspection at real cadence is physically impossible, and sampling leaves gaps by definition. There is something worse than fatigue: the criteria. What counts as an acceptable wrinkle changes between fabric and leather, between colors, between operators, and it degrades across the shift. There is no objective standard, there are several people doing their best. And in a sequenced operation the defect does not stay home. It goes on the truck in the exact order, reaches the customer line and is found there, where it costs an order of magnitude more and where it turns into quality points and into your supplier rating.

• Wrinkles in the trim cover, an open seam or a skipped stitch, a missing clip or hog ring, uneven leather shade between backrest and cushion, a mislaid label: trim quality is decided by eye.
• One hundred percent visual inspection at real cadence, with variants changing every few units, is physically impossible, and sampling leaves gaps by definition.
• What counts as an acceptable wrinkle changes between fabric and leather, between colors, between operators, and degrades across the shift. There is no objective standard, there are several people doing their best.
• In a sequenced operation the defect does not stay home: it goes on the truck in the exact order and is found on the customer line, where it costs an order of magnitude more and turns into quality points and your supplier rating.

How it fits the IRIS system

Edge — local inference in milliseconds, trained on your own variants, the image never leaves the plant.

Edge iLEAN puts a camera over the trimming station, with local inference in milliseconds. The model is trained on good and bad examples of your plant's real variants, not on a generic catalog. And inference runs locally, on the unit next to the camera. No cloud latency, no cost per image, no bandwidth. And above all: the image never leaves the plant. The flagged unit is diverted to rework before entering sequence, not after. From one or two percent sampling to one hundred percent inspection, with a single auditable criterion that does not depend on the time of the shift.

The flagged seat is diverted to rework before entering sequence, not after. From one or two percent sampling to one hundred percent inspection, with a single auditable criterion that does not depend on who is on shift or what hour it is. The wrinkle stays home, and so do the quality points it would have cost.

See the full IRIS architecture →

Before and after

Today's trim inspection versus Edge vision

Aspect	Visual inspection	With iLEAN Edge
Coverage	1-2% sampling	100% of seats
Acceptable wrinkle	Depends on operator, color and hour	One auditable criterion
Missing hog ring or clip	Found on the customer line	Found at the trimming station
Leather shade between backrest and cushion	Judged by eye	Compared against the variant's reference
Where the defect is caught	On the truck, in sequence	Before entering sequence
Image and inference	—	Local, no cloud, no cost per image

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 rework ratio and the cost of quality points with the customer, to validate against your claims history.
• From 1-2% sampling to 100% inspection with one criterion that does not drift across the shift.
• Fewer trim defects reaching the customer line, which is where they cost an order of magnitude more and turn into quality points.
• And the return that never shows in the accounts: your position in the customer's quality ranking, which decides the award of the next program and is where the rework ratio is really measured.

Estimated payback 5-12 months · from 1-2% sampling to 100% inspection Estimated payback of five to twelve months depending on your current rework ratio and the cost of quality points with the customer, an estimate to validate against your claims history. The bigger return never shows up in the accounts: it is your position in the customer's quality ranking, which is what decides the award of the next program.

And the fair question from the production manager

“What if it sends good seats to rework?” — the false positive is the real risk of any vision system, which is why the model is trained on good and bad examples of your plant's real variants and lighting, not on a generic catalog. Judging a cover against its own variant reference is an anchored task where the best models drop below 1.5% error [1]. And borderline cases are not simply rejected: they are escalated for a person to decide, and every decision feeds training.

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