AI vision poka-yoke — human assembly errors are not prevented by a poster.

The problem

Human error in assembly does not belong to the person who makes it — it belongs to the system that allows it.

Poka-yoke (mistake-proofing) is one of the pillars of jidoka: making it physically impossible for an error to reach the next step. It works perfectly when you can design a mechanical fit that only goes in one way. But in multi-variant manual assembly, where every SKU has a different BOM and the same station builds version A today and version C tomorrow, mechanical poka-yoke becomes impossible or prohibitively expensive — and what lands on the operator instead is posters, checklists and Monday-morning training.

And the operator does it right. 99.8% of the time. But the remaining 0.2% is what reaches final inspection, the customer, or (worse) the field. In automotive the demanding standard is in the order of 25 PPM — every defect that slips through costs hard money and, above all, customer trust.

The acute problem shows up at the changes: SKU change, shift change, new operator. That is precisely when the poster on the wall is not enough. The classic answer is to add a downstream check, a rework loop, or an inspector. It is expensive and, worse, it arrives late — the defect has already been born.

How it fits the IRIS system

iLEAN does not replace the operator — it gives them a safety net before the next step.

Vision poka-yoke is jidoka in two stages: first it detects the error at the moment it is born; then (as data accumulates) it anticipates the root cause so the error stops happening at all. The filler iLEAN brings seals the gap between the veteran operator's expertise and the new operator's risk, without asking you to redesign the workstation or add an inspector on the line.

The camera sees the step, the agent compares it with the sequence for that SKU, the operator gets the warning before moving on. The part does not slip through with the fault — and the person signs off when they have corrected it.

The iLEAN pieces applied to vision poka-yoke:

• iLEAN Vision (on Edge) — a physical terminal with a camera and a convolutional neural network (CNN) that sees every assembly step in milliseconds. It works without a network: if the plant loses WiFi, Edge keeps seeing, comparing and warning, because what is critical cannot depend on connectivity.
• Validated sequence per SKU — the system first reads the SKU being built (label, code, configuration) and loads the expected sequence. In multi-variant work this is what removes the “I built variant A with variant B's BOM” error.
• Feedback to the operator, not to the inspector — the warning goes to the station: green/red light, watch vibration, tone in the earpiece. The operator corrects where the error is born, not after the part has moved to the next station. The person signs off once they have corrected it and the part can move on.
• Capturing the veteran's know-how — training the sequence with the operator who performs it best leaves that knowledge as a permanent capability of the plant. The day they retire, the sequence lives on in the camera.

See the full IRIS architecture →

Before and after

Classical poka-yoke vs. AI vision poka-yoke

Aspect	Classical poka-yoke (poster + checklist)	With iLEAN Vision + Edge
Error detection	At final inspection or customer claim	At the station itself, before the next step
Multi-variant	Confusion at SKU changeover	Reads the SKU and loads the correct sequence
New operator	Learning curve paid in scrap	Safety net from day one
Shift handover	Context lost at the relay	Same criterion across every shift
Veteran's know-how	Walks out the day they retire	Captured as a permanent capability
Operation without network	n/a	Edge keeps running on panel power

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 set it out so the committee has an order of magnitude; we refine it during the diagnostic.

• Manual assembly line with 3–8 SKUs and different BOMs, a recurring defect caught at final inspection that originates at one specific station.
• Vision pilot at the critical station: camera, network trained with the veteran operator, local feedback to the operator. First value expected within a few weeks: the camera starts catching the first family of errors.
• Reduction of the defect generated at that station estimated at ≥ 30% over the first months. Indicative payback between 4 and 9 months, depending on the average cost of the defect and your current PPM.
• In automotive the demanding standard in the order of 25 PPM ^[2] is the hard lever: every PPM avoided is hard cost and, above all, reduced customer risk.

And the operator's reasonable doubt

“If the camera corrects me at every step, I'm going to assemble slower.” It is the right objection — and the answer is not theoretical, it is a design choice: feedback goes to the operator, not to the supervisor; it stays green by default and only switches to a warning if there is a deviation; and the integration with the station is built so as not to break the flow. What changes for the operator is that they start the shift with peace of mind — the camera is a net, not a whip. The warm bed: the previous shift went well because the camera secured every step.

And for the CAIO: hallucination is a problem of free generation, not of anchored tasks. Comparing an image against a validated sequence is the most anchored task there is — the best models brought the error below 1.5% ^[1]. And even so, the part does not move on by itself: the person signs off once they have corrected it.

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

^[2] Symestic — PPM quality standards in automotive.