One hundred percent appearance inspection on interior parts

The problem

The defect catalog is long, and the OEM defines it.

Sink marks, splay, flow lines, ejector marks, burns, specks or dirt inclusions in the paint, uneven gloss, a tone mismatch between the part and the assembly. The defect catalog for an interior part is long and allows no aesthetic debate: the customer defines it. One hundred percent inspection by eye is impossible at real line rate, and sampling leaves gaps exactly where intermittent defects live: the ones from one specific cavity, the ones that appear after a purge, the ones that only show up once the machine has been hot for two hours. And there is an added criterion problem: what mornings rejects, nights lets through. The argument with the customer about acceptance limits is fought without data, and the customer always wins it.

• Sink marks, splay, flow lines, ejector marks, burns, specks or dirt inclusions in the paint, uneven gloss, a tone mismatch between the panel and the rest of the module. None of it is negotiable: the OEM's boundary sample defines what is acceptable.
• One hundred percent inspection by eye is impossible at line rate, and sampling leaves gaps exactly where the intermittent defects live: one cavity, the parts after a purge, the ones that appear once the mold has been hot for two hours.
• What mornings rejects, nights lets through. The argument with the OEM about acceptance limits is fought without data, and the OEM always wins it.
• And an appearance part that reaches the vehicle maker is not scrap: it is a complaint, one hundred percent rework and, if it repeats, controlled shipping with an external inspector paid for by the plant.

How it fits the IRIS system

Edge at the press exit and the paint line exit — inference on a local GPU, in milliseconds.

Edge, with an industrial camera at the press exit and another at the paint line exit, diffuse and dark-field lighting, and a convolutional neural network on a local GPU trained on good and bad parts of the plant's actual part numbers and on the acceptance limits agreed with the customer. Inference is local: neither latency, nor cost, nor bandwidth allow images to be sent to the cloud at that rate. The defective part is separated before it enters module assembly, where recovering it already costs ten times more. Every reject is recorded with its image, its part number, its mold and its cavity.

The defective part is separated before it enters module assembly, where recovering it already costs ten times more. And every reject is stored with its image, its part number, its mold and its cavity: the argument about limits finally has data on both sides. Across three shifts the criterion is the same one, trained on the OEM's boundary samples.

See the full IRIS architecture →

Before and after

Visual inspection versus Edge vision on interior parts

Aspect	Visual inspection	With iLEAN Edge
Coverage	1-2% sampling	100% of parts
Acceptance criterion	Drifts between shifts	One criterion, trained on the boundary samples
Intermittent defect from one cavity	Missed by sampling	Reject history per cavity
Point of detection	At the OEM, or in assembly	Before module assembly
Where the images go	—	Nowhere: local GPU
Dispute over limits	Opinion against opinion	Image and record per reject

from 1-2% sampling to 100% inspection · from a criterion that drifts between shifts to a single traceable criterion · from suspecting a cavity causes trouble to a reject history per cavity that tells you when to work on the mold.

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 current scrap and rework rates.
• Fewer OEM appearance complaints, which are the ones that end in one hundred percent rework and containment.
• Lower containment cost: no external inspector paid for by the plant on controlled shipping.
• And a reject history per cavity that tells you when to work on the mold, instead of suspecting it.

estimated payback 5-12 months depending on current scrap and rework rates; fewer customer appearance complaints and lower containment cost with external inspection. *Estimate to be validated.*

And the fair question from the production manager

"What if it rejects good panels?" — the false positive is the real risk of any vision system, which is why the network is trained on good and bad parts of your own part numbers, under that station's lighting, and on the acceptance limits agreed with the OEM, not on a generic model. Borderline parts are not silently rejected: they are escalated to a person. And where the system reads a label or a mold number rather than a surface, that is an anchored task where the best models drop below 1.5% error [1].

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