AI control of vehicle rear camera modules — the module passes the functional test, fails in the field, and the OEM returns the whole batch.

The problem

The functional test says OK — and the module fails in the field months later.

The quality manager of a camera module line knows the pattern well: the module passes the electrical test at end of line — power, communication, video signal — and leaves the plant packed with its OK. Months later the claim arrives: the image is off in the parking guide lines, there is condensation inside the optics. The module that passed was not right — it was inside what the functional test can see, which is not everything the module is.

Three things come together in a rear camera module, and almost no end of line covers all three at once:

1. The optics are not verified piece by piece — focus and optical axis alignment are set at the station, but a fine misalignment the electrical bench cannot tell apart ends up on the customer's screen. The module works; it works crooked.
2. Watertightness is decided before closing, and after that nobody sees it — the module is exposed to the weather; a sealant bead with a narrow or displaced stretch passes the day-one functional test and lets moisture in months later. Once the module is closed, no inspection is possible.
3. Traceability by VIN is rebuilt by hand — when the OEM raises a claim, cross-referencing the module's serial number with its process parameters and the vehicle's VIN means days of spreadsheets. In the meantime, the OEM's prudent answer is to return the whole batch.

The result is always the same: the plant pays for batch returns where most modules were fine, the field claim is answered late and without data, and the knowledge that “this sealing station goes off-center when…” lives in the head of the veteran setter. The day they leave, it leaves with them.

How it fits the IRIS system

iLEAN does not replace your end of line — it adds the eyes and the memory it lacks.

The problem with the camera module is not that a test is missing: it is that the test you have only covers the electronics, and what fails in the field is the optics and the watertightness. iLEAN completes the end of line with 100% vision verification, ties every result to each module's serial number and to the destination VIN, and holds the flow when the process drifts — without changing the test bench you already have.

Edge verifies optics and sealing with vision at end of line. Connect integrates the electrical test and ties the result to the serial number and the VIN. JIDOKA AI holds the flow on drift. The person decides what gets released.

The iLEAN pieces applied to rear camera module control:

• Edge — a local terminal with vision at end of line that checks, on 100% of modules, the alignment of the optical axis against a calibrated reference and the quality of the sealant bead (continuity, width, position) before closing, which is the last moment the seal can be seen. It runs locally: if the plant loses the network, Edge keeps verifying, recording and holding whatever does not pass. What is critical does not depend on WiFi.
• Connect — the integration layer that hooks into your existing electrical test bench and ties its result, together with Edge's checks and the sealing station parameters, to each module's serial number; and when the module is assigned to a vehicle, to the destination VIN. A single record per module, built on the line, with no manual transcription.
• JIDOKA AI — holds the flow on process drift, not only on the one-off failure: bead width falling while still inside tolerance, average misalignment growing shift after shift, electrical test spread opening up. It holds the affected modules and alerts with the specific signal; the supervisor decides whether to release, rework or adjust the station.
• Agents — when a field claim arrives, they correlate the reported symptom with the process signature of that specific module (its bead, its alignment, its test) and with the modules that share that signature. Exposure is scoped to the modules with a common cause, instead of assuming the whole batch.

See the full IRIS architecture →

Before and after

Classic end of line vs. end of line with iLEAN

Aspect	EOL with functional test	With iLEAN Edge + Connect + JIDOKA AI
Optical axis alignment	Set at the station, no 100% verification	100% vision against a calibrated reference
Sealant bead	Operator's eye, invisible once closed	Vision before closing: continuity, width, position
Electrical test	Result isolated in the bench	Integrated and tied to the module's serial number
Process drift	Discovered once failures are already there	JIDOKA AI holds the flow on the trend
Field claim	Days rebuilding records by hand	Hours, with the exact module's signature by VIN
Batch return	Whole batch, because it cannot be scoped	Scoped to the modules with a common cause

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.

• Rear camera module line with a functional test bench at end of line, a sealing station whose parameters are logged locally, VIN traceability rebuilt by hand and at least one documented batch return in the last year.
• Edge pilot with vision on alignment and the sealant bead + Connect tying your existing electrical test to the serial number. First value expected within a few weeks: the single record per module and the first drift holds show up early, before the learned pattern is fully tuned.
• Expected reduction in batch returns from avoidable causes — fine misalignment and marginal sealing, both detectable at end of line — of ≥30% in the first few months, a defensible floor and always an estimate to be validated.
• Field claims answered in hours with data on the exact module, instead of days of manual reconstruction — and exposure scoped to the modules that share the signature, not to the whole batch.
• Indicative payback between 5 and 12 months. The hard lever: every batch the OEM does not return is freight, re-inspection and penalties wiped out at once.
• A recurring benefit that does not enter the ROI but weighs: the veteran setter's judgement about the sealing station becomes a permanent capability of the plant, not of the person who is about to retire.

And the quality manager's reasonable doubt

“What if the AI correlates a claim wrongly and I hold good modules?” — the iLEAN system never decides material disposition on its own. JIDOKA AI holds and proposes; the quality manager decides whether to release or rework. Hallucination is a problem of free generation, not of anchored tasks: in tasks where the AI cross-references a module's process signature with its test result and its history, the best models brought error below 1.5%^[1]. And even so, what is critical goes through the safety rings — the Agents live in the outer ring, propose inward, and batch disposition is signed by a person. Never the other way round.

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