Critical seams on the EN 361 fall arrest harness with vision AI — the defect that eye fatigue lets through is the one that ends up in court.

The problem

The sustained human eye on a line does not exist — the expert report does.

An EN 361 fall arrest harness is made from polyester webbing, metal buckles and seams of Dyneema or high-tenacity nylon thread stitched in reinforced patterns over the critical points: dorsal D-ring, sternal D-ring, chest buckle joint, reinforcement at the attachment points. The piece has to hold a real human fall. And the defects that compromise it are:

1. Skipped stitches in the reinforcement of the attachment point — millimeters that unzip the seam when the dynamic load arrives.
2. Irregular thread tension — the stitch looks well made until it is pulled and comes loose.
3. Deviation from the programmed pattern — the seam closes outside the intended reinforcement, on free webbing.
4. Mixed thread spools — an undocumented thread change halfway through the batch.

Sustained 100% human control, shift after shift, is not real: the eye tires, the pace is unforgiving, and subtle defects slip through on 1-2% of production. In automotive, the demanding standard is in the order of 25 PPM (parts per million) ^[1]; in fall protection PPE, where one piece can end up in a court expert report, that 1-2% is not an acceptable number — it is a brand risk and a legal defense risk.

How it fits the IRIS system

iLEAN Vision does not replace the sewing operator — it puts a net underneath.

The problem with critical seams is not that the operator is careless: it is that sustained human fatigue is a constant of reality, and nobody can look at 100% of the pieces with the same attention at eight in the morning as at eight in the evening. iLEAN acts as the putty that fills the gap between the sewing operator's skill and the inevitable fatigue of the shift, without asking the operator to be infallible.

Vision sees every critical seam on the line. The agent pairs it with the serial number. The harness file is assembled. The person validates — the safety net is held up by the machine.

The iLEAN pieces applied to critical seams on the EN 361 harness:

• Vision (an Edge with a CNN trained on EN 361 seams) — installed at the exit of every critical sewing machine. It looks at every dorsal D-ring, every buckle joint, every attachment point reinforcement. It detects skipped stitches, loose stitching, pattern deviation and changes in thread tension. If it finds a defect: stack light + ejector in milliseconds — before the serial number is marked. It works with no network: if the plant loses WiFi, Vision keeps looking.
• Connect — captures spool changes, incident reports, the pattern recipe from the MES and the buyer's thread certificates. Everything hangs off the batch.
• Traceability agent — links the image of the critical seam with the serial number about to be marked on the piece and files the pair under the batch record in Central. If in three years an expert report arrives for the harness with serial number XYZ, the dossier comes out in minutes.

And the safety boundary: Vision holds; the quality manager signs. The three rings of the IRIS system guarantee that no piece flagged as defective by the AI goes back onto the line without human validation.

See the full IRIS architecture →

Before and after

Human visual control vs. critical seams with iLEAN Vision

Aspect	Human visual control	With iLEAN Vision + agent
Inspection coverage	Sampling (1 in 10) or 100% with accumulated fatigue	Genuinely 100%, held up by the camera shift after shift
Skipped stitch detection	1-2% get through, depending on pace and shift	Detected and ejected in milliseconds
Undocumented spool change	The operator remembers, or does not	Captured by Connect, linked to the batch
Link to the harness serial number	Nonexistent — you trust the piece is good	Photo of the critical seam + serial number in Central
Expert report three years later	“We don't have the detail for that harness”	The file for that specific harness in minutes
Recall of a model from the market	Precautionary, the whole batch just in case	Limited to the pieces flagged on the line

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.

• Fall protection PPE plant with 3-4 critical sewing lines, multi-model (EN 361, EN 358, EN 813), traceability by serial number.
• Vision pilot on one line (camera over the critical seams + actuator + integration with the MES and serial number marking). First value within a few weeks.
• Indicative payback between 4 and 9 months, depending on volume and the frequency of incidents documented over recent years. Expected reduction of escaped defects ≥ 30% from the first pilot line onward.
• The hard lever is a single legal claim avoided and the chain of evidence available for defense in an expert report. One serious claim pays for the entire installation.

And the quality director's reasonable doubt

“What if Vision flags good pieces as bad?” — hallucination is a problem of free generation, not of anchored tasks. Vision does nothing but compare the real seam against the conforming pattern — it recontextualizes an image against an approved pattern. In anchored tasks, the best models brought error below 1.5% ^[2]. And even so, no flagged piece goes back onto the line without the operator validating it and the quality manager signing.

^[1] Automotive quality standard ≈ 25 PPM (parts per million). Symestic.
^[2] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in tasks anchored to the source.