The cabinet defect, caught before it reaches the customer

The problem

The cabinet is the part the bottler sees, and the least inspected.

At line cadence, the human eye gets tired and lets through sheet-metal dents, paint defects, or a glass door seal that doesn't close tight. The defect reaches the end customer's point of sale — visible, brand-facing, expensive to fix in the field.

• At line cadence, the inspector's eye gets tired. Sheet-metal dents, paint runs and a glass door gasket that does not seal flush slip through. The inspector is usually alone at the end of the line, looking at every cabinet at the same pace for the whole shift.
• The cooler is a brand-facing object: it ends up at the bottler's or the retailer's point of sale, in front of their shoppers.
• A defect there is visible, embarrassing for the customer and expensive to fix in the field, often with a technician traveling to the store. The bottler or retailer judges the plant by what its shoppers see in the store, not by the yield report.
• Door seal problems are worse than cosmetic: a gasket that does not close tight raises power draw and makes the cooler fail the customer's own checks.

How it fits the IRIS system

Edge — a camera over the line and a model trained on that cooler model's cabinets.

An Edge camera over the line with a neural network trained on good and bad cabinets of that specific model catches the defect in milliseconds and pulls the unit before packing.

The camera does not get tired at the end of the shift and inspects the hundredth cabinet like the first. The unit with the defect is pulled before it is boxed, which is the last point where fixing it is cheap. It also gives Quality something it does not have today: a count of each defect type per model, per shift and per station.

See the full IRIS architecture →

Before and after

Cabinet inspection by eye versus Edge vision

Aspect	Visual inspection	With iLEAN Edge
Coverage at line speed	Depends on fatigue	Every cabinet, every shift
Sheet-metal dent	Seen if the light is right	Detected in milliseconds
Paint run or scratch	Inconsistent	Classified per unit
Door gasket not seating	Rarely checked	Checked on every door
Where the defect is found	At the point of sale	Before packing
Field repair at the store	Recurring cost	Avoided

Human visual inspection at line speed → consistent Edge inspection shift after shift.

Impact estimate

Impact estimate — to be validated with your numbers.

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 5-10 months, depending on line speed and the current reject rate.
• Cabinet defects rejected before packing, not discovered at the bottler's or retailer's point of sale. Every defect stopped at the plant is a field visit avoided at the customer's store.
• Human visual inspection at line speed becomes consistent Edge inspection shift after shift. The defect statistics per model show which upstream station is producing them.
• And the inspector moves from looking at every cabinet to deciding on the borderline ones the model escalates.

Estimated payback of 5-10 months depending on line speed and current reject rate. Estimate to validate.

And the fair question from the production manager

“A paint shade or a reflection on the glass door will make it reject good coolers.” — that is the real risk, which is why the model is trained on good and bad cabinets of your specific models, under your line's lighting. Classifying a known defect type on a fixed product is an anchored task, where the best models drop below 1.5% error [1]; the borderline cases go to a person, and each decision improves the model.

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