The stitching is what holds the pressure

The problem

The defect that generates no signal until the day of the crash.

Stitch inspection is typically visual and by sampling, or at 100% but depending on one person's sustained attention at machine pace for eight hours. Modern machines detect thread break, but they do not detect a skipped stitch or a drift off pattern, which are precisely the failure modes that count. And there is a very short time window: once the part is folded inside the module, the stitching can no longer be inspected. Any later doubt means destroying the unit to look at it. That is why the control point has to be before folding, or it does not exist. What makes this defect unique is that it generates no signal: no complaint, no return, no quality indicator that moves. The failure shows itself exactly once, on the day it actually mattered.

• When the gas generator fires, what stops the cushion from bursting is the stitching. A skipped stitch, a thread at uneven tension or a seam drifting off pattern are not visible to the naked eye, and at the machine's real cadence the human eye cannot keep up.
• Inspection is typically visual and by sampling, or at 100% but depending on one person's sustained attention at machine pace for eight hours. Modern sewing machines detect thread break, but not a skipped stitch or a drift off pattern, which are precisely the failure modes that count.
• The time window is very short: once the cushion is folded inside the module, the stitching can no longer be inspected. Any later doubt means destroying the unit to look at it. The control point has to be before folding, or it does not exist.
• What makes this defect unique is that it generates no signal: no complaint, no return, no quality indicator that moves. It shows itself exactly once, on the day it actually mattered.

How it fits the IRIS system

Edge — local CNN vision at the sewing output, before folding.

Edge — local CNN vision on the line.

An industrial camera over the output of the sewing machines and a local Edge GPU: inference in milliseconds, cushion by cushion, on a network trained on good and bad examples of each variant's stitch pattern. The defective part is separated before it enters folding, and the image of every part is tied to its record, so the visual evidence exists even once the cushion is folded and can never be seen again.

• Industrial camera over the output of the sewing machines.
• Local Edge GPU: inference in milliseconds, part by part. No image leaves for the cloud, for latency, for cost, and because the cushion design is the customer's property.
• Network trained on good and bad examples of each manufactured variant's stitch pattern.
• The defective part is separated before it enters folding.
• The image of every part is tied to its record, so the visual evidence exists even once the part is folded.

See the full IRIS architecture →

Before and after

Today's seam inspection versus Edge vision

Aspect	Visual inspection	With iLEAN Edge
Coverage	Sampling, or attention that fades	100% of cushions, at cadence
Skipped stitch and pattern drift	Invisible to the eye at machine pace	Classified part by part
Point of detection	Downstream, or never	Before folding, while it is still visible
Proof that a seam was good	None once the cushion is folded	An archived image per unit
Where the images go	—	Nowhere: local GPU, nothing leaves the plant
New cushion variant	Retrain the inspector's eye	Add its stitch pattern to the network

from sampling (or sustained human attention) to 100% of parts controlled · from defects found downstream — or never — to detection where it is still visible and still cheap · from "we cannot prove that seam was good" to an archived image per unit.

Impact estimate

Impact estimate — to be validated with the quality manager.

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 cadence and current defect rate.
• The dominant value is not the scrap avoided: it is eliminating a failure mode that otherwise only shows up in the crash.
• From sampling — or sustained human attention — to 100% of cushions controlled where it is still visible and still cheap.
• From "we cannot prove that seam was good" to an archived image per unit, ready for the containment team or the carmaker.

estimated payback 5-12 months depending on cadence and current defect rate. The dominant value is not the scrap avoided but eliminating a failure mode that otherwise only shows up in the crash. *Estimate to validate with the quality manager.*

And the fair question from the production manager

"What if it rejects good cushions and stops the sewing line?" — the false positive is the real risk of any vision system, which is why the network is trained on good and bad seams of your specific variants, with that station's lighting, not on a generic model. Classification against a known stitch pattern is an anchored task, where the best models drop below 1.5% error [1]; borderline cases are not simply rejected, they are escalated to a person, and every decision feeds back into training.

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