Car upholstery with AI — the shade and the stitching are caught the first time, or they are caught at the dealership.

The problem

The line forgives an upholstery defect — the end customer does not.

Upholstery is the cosmetic part of the interior par excellence: what the customer touches every day and what they look at under the overhead light of the dealership. And yet it is one of the few operations in an automotive Tier 1 where the main filter is still the veteran operator's eye.

1. The material is not homogeneous — every batch of leather or fabric has its own shade, its own grain, its own behavior under light. Two panels of the same color loaded from different batches can clash in the finished car.
2. Sewing is manual or semi-automatic — and the skipped stitch, the puckering or the pattern deviation show up in small but recurring numbers, above all at the end of the shift.
3. Gaps when assembling onto the frame appear downstream — when the panel is already sewn, bonded and almost on the truck.
4. The field claim is the worst — a seat returned from a dealership for a cosmetic defect costs ten times what it costs to recover it in cell 2.

The traditional filter — veteran operator + final review + field claim — works, but it lets through whatever a tired eye at the end of the shift lets through. And the end customer forgives none of them.

How it fits the IRIS system

iLEAN does not replace the veteran operator — it captures their eye and sustains it shift after shift.

The upholstery problem is not a lack of information: it is information that lives in the veteran's eye and walks out the door the day they retire. iLEAN is the putty that captures that pattern and leaves it as a permanent capability of the line.

Edge sees the panel and the stitching. Connect reads the material batch. The agent cross-references shade, stitching and batch with the acceptance history and raises the alarm before the next cell. The person decides.

The three iLEAN pieces applied to car upholstery:

• Edge — cameras with controlled lighting over the sewing area and over the finished panel. CNNs trained on samples from your own plant recognize material defects (stain, scratch, shade irregularity), stitching defects (skipped stitch, loose thread, puckering, pattern deviation) and die-cutting defects. Verdict in milliseconds, without stopping the line. It works with no network.
• Connect — captures the batch of every roll of leather, fabric or Alcantara that enters the line (whether by digitizing the delivery note, reading a label, or integrating with the warehouse system). It also captures what arrives from outside: the supplier's email flagging a batch change, the SQA notice about a field claim from last week.
• Agent — cross-references shade and stitching with the batch, with that supplier's acceptance history and with the SQA claims. It does not just detect a one-off defect: it detects a batch trend before the panel reaches the dealership. It alerts the quality manager. The person decides whether to accept, to scrap, or to change the cutting direction to minimize the defect.

See the full IRIS architecture →

Before and after

Visual inspection + field claim vs. AI vision + batch trend

Aspect	Operator's eye + final review	With iLEAN Edge + Connect + Agent
Moment of detection	End of shift or field claim	Cell exit, in milliseconds
Shade between batches	Verbal, intuition, no traceability	Learned envelope + cross-check with supplier history
Defective stitching	Caught in the final review or never	Caught step by step, with a rework flag
Field claim	High cost, brand damage	Early warning: suspect batch held before it ships
Capturing the veteran's eye	Walks out with the veteran the day they retire	Stays as a trained model, permanently
Operation with no network	n/a	Edge keeps running on the light from the panel

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Plant with automatic cutting + semi-automatic sewing + assembly onto foam/frame, a mix of leather / fabric / Alcantara, at least 20 part numbers in flow.
• Edge pilot in 2 critical cells (post-sewing review + final review before bonding), cameras with controlled lighting and training on your own samples. First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the current frequency of field claims for cosmetic defects and the average cost of rework.
• Hard levers: rework avoided on an already sewn part + field claim avoided (the most expensive one) + a file per seat number for the OEM.
• The automotive quality standard is on the order of 25 PPM ^[1]: with upholstery in plain sight of the end customer, the real margin is even tighter.

And the interior quality manager's reasonable doubt

“What if the AI mistakes a shadow from the lighting for a defect and scraps a good part?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI classifies an image against patterns trained on your own samples, the best models brought error below 1.5% ^[2]. And even so, what is critical is never decided alone: the system flags the part for review, the person validates. iLEAN's three safety rings exist precisely for this.

^[1] Symestic — automotive quality standard ~25 PPM.

^[2] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.

Related processes: automotive seat manufacturing · body panel sheet metal defects · digitized OEM CSR.