The defect gets caught before the cabinet closes, not in the field

The problem

Once the cabinet is closed, a wiring defect becomes a data-center problem.

100% human visual inspection is inconsistent, even more so with new hires joining at a fast pace. If the defect makes it inside a closed cabinet, it can turn into a field failure inside a data center — extremely costly to diagnose and repair on site.

• Before the doors and covers go on, someone looks inside the cabinet: wiring routes, bus joints, barrier placement, breaker seating, torque marks on the bolts. 100% human inspection, at line pace.
• The eye tires, and with new hires joining every week it is even less consistent. A crossed control wire or a loose busbar connection gets through late in the shift, on exactly the unit nobody remembers inspecting.
• If it reaches the FAT, it costs rework on a closed, finished unit and a finding in front of the customer. If it gets past the FAT, it becomes a field failure inside a data center.
• Diagnosing and repairing on site, under the customer's outage and access rules, costs many times what catching it at the closing station would have. And it is the kind of incident a hyperscale customer remembers when awarding the next project.

How it fits the IRIS system

Edge — a camera over the closing station and a model trained on your own cabinets.

Edge — local CNN vision, no image sent to the cloud (latency, cost, customer design confidentiality). Model trained on correct and incorrect examples of the line's specific product.

Inference runs on the line, in milliseconds, and no image of the customer's design leaves the plant. The camera does not replace the technician's eye; it makes sure the eye is not the only barrier at the last moment the interior can be seen. After the door closes, nobody looks again until the customer does.

See the full IRIS architecture →

Before and after

The closing inspection today versus Edge vision at the closing point

Aspect	Human inspection	With iLEAN Edge
Cabinet interiors inspected	Sampled in practice, by fatigue	Every unit at the closing point
Crossed control wire	Missed late in the shift	Detected before the doors go on
Loose or missing bus joint hardware	Depends on who looks	Classified unit by unit
Where a cabinet wiring defect is found	At FAT, or in the field	At the closing station
Images of the customer's design	—	Processed locally, never sent out
Unclear cases	Passed or argued over	Escalated to the quality lead

Inconsistent visual sampling → 100% unit inspection at the closing point.

Impact estimate

Pre-closing vision: estimated impact, to validate with your rework rate.

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 the current rework rate at FAT and after closing.
• Fewer FAT findings and less late-stage rework on closed, finished units, where every hour of rework is the most expensive of the build because covers, barriers and doors have to come off again.
• Inspection goes from inconsistent visual sampling to every unit at the closing point, with an image of each interior kept for the FAT dossier.
• And the defect that would have turned into an on-site repair in a data center stays inside the plant, where fixing it takes a screwdriver and ten minutes instead of a service trip.

Estimated payback 5-12 months depending on current rework rate, reducing FAT findings and late-stage rework. *Estimate to validate.*

And the fair question from the production manager

“What if it stops good cabinets and slows the line?” — false positives are the real risk of any vision system, so the model is trained on correct and incorrect examples of your own product, under that station's lighting. Checking presence and position of known parts is an anchored task, where the best models drop below 1.5% error [1], and unclear cases go to a person instead of stopping the line silently.

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