The Edge camera that inspects 100% of your garments

The problem

At line cadence, the eye that inspects is the first thing to tire.

100% human visual control is impossible at real line cadence, especially on long shifts. Sample-based inspection leaves gaps.

• A crooked seam, a skipped stitch, a bad buttonhole, a stain or a misplaced reflective tape: each is small, and each is a program complaint. On high-visibility garments, a tape off position can also take the garment out of its EN ISO 20471 class.
• 100% human visual control is impossible at real line cadence, especially at the end of a long shift. The inspector who is sharp at the start of the shift is not the same one eight hours later.
• So plants sample 1-2% of garments, and everything between samples ships on trust. On a uniform program with thousands of identical garments, that trust is exactly what the customer is paying for.
• Second-quality garments found by the customer cost far more than the ones found at the inspection table. They cost the garment, the shipping, the replacement and part of the customer's trust.

How it fits the IRIS system

Edge — local vision trained on your own styles, no cloud dependency.

Edge — local CNN vision, no cloud dependency. The model is trained on good/bad examples of the plant's actual style. Defective garments are pulled before folding and packing.

The model is trained on good and bad examples of the plant's actual styles, not on a generic garment. A navy coverall and a high-visibility jacket are learned as what they are. Defective garments are pulled before folding and packing, with the defect type and the image attached, which is the last point where pulling one is still cheap. After that point, every defect costs a carton opened, a program complaint or a return.

See the full IRIS architecture →

Before and after

Sampled inspection versus Edge vision

Aspect	Visual inspection	With iLEAN Edge
Coverage	1-2% sample	100% of garments
Late-shift fatigue	Lets borderline defects through	Constant at cadence
Where the defect is caught	At the customer	Before folding and packing
Inference	—	Local, milliseconds per garment
Borderline seam	Judged in a hurry	Escalated to the inspector
Defect record	A tally on paper	Per garment, style and line

1-2% sampling → 100% garment coverage. Program complaints for visual defects → drastically reduced.

Impact estimate

Impact estimate — to be validated against your current second-quality ratio.

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 current second-quality ratio.
• Coverage goes from a 1-2% sample to 100% of garments. Not a bigger sample: every single garment that reaches the inspection table.
• Program complaints for visual defects drastically reduced. Customer programs measure suppliers on exactly those complaints.
• And defects are recorded per style and line, which tells sewing where they originate. That record is what lets a line lead fix the operation instead of only rejecting the garment.

estimated 5-12 month payback depending on the current second-quality ratio. *Estimate to validate*.

And the fair question from the production manager

“What if it pulls good garments?” — the false reject is the real risk of any vision system, which is why the model is trained on your own styles and that table's lighting. Classifying known defect types on a known garment is an anchored task, because the catalog of defects for each style is short and repetitive, where the best models drop below 1.5% error [1]. And borderline garments are not discarded: they go to the inspector, and every decision feeds training. The inspector keeps the last word on every garment the model is not sure about.

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