A leaking container is not a return: it is an in-transit spill

The problem

A defective chemical container does not come back: it spills in transit.

A defective chemical container does not produce an ordinary return. It produces a spill inside a hazardous goods truck, a trailer clean-up, a safety notification and, depending on the product, an incident reportable to the authorities. The receiving customer does not reject the container: they reject the whole shipment. Today's inspection depends on the attention of a person doing three other things. It is not a matter of professionalism: at sustained cadence the human eye fails, and it fails precisely after hours of watching. The defect that slips through is caught at the branch or at the customer. That is, late, far away and with witnesses.

• A cap not fully seated, a neck that weeps a few drops, a seal that did not bond, a label that is crooked or unreadable. Each one gets through the human eye at line cadence, and each one becomes a spill inside a hazmat truck.
• The consequence is not a return. It is a trailer clean-up, a safety notification, an incident reportable to the authorities depending on the product, and a customer that rejects the whole shipment, not the one drum.
• Today's inspection depends on the attention of an operator who is also filling, labeling and palletizing. At sustained cadence the human eye fails, and it fails precisely after hours of watching.
• The defect that slips through is caught at the branch or at the customer: late, far away, and with witnesses.

How it fits the IRIS system

Edge — AI vision over the line exit, local inference, before wrapping.

Edge, AI vision on the line.

The model is trained on the plant's own containers — its jugs, its drums, its IBCs, with their caps and their labels — not on a generic catalog. Inference runs locally in milliseconds, the image never leaves the plant, and the rejected container is pulled before the pallet is wrapped, with the photo stored against the batch. The operator stays on the line; the camera takes over the part of the job the human eye was never going to do reliably for eight hours.

• Industrial camera over the packaging line exit, before stretch-wrapping.
• Vision model trained on images of the plant's real containers — jugs, drums and totes, with their caps and their labels — not on a generic catalog.
• Local inference in milliseconds: the data never leaves the plant.
• Detects incomplete closure, moisture or weeping at the neck, a missing, crooked or unreadable label, and deformed containers.
• Flags or rejects the container before wrapping and stores the image linked to the batch, where it later serves as auditable evidence.

See the full IRIS architecture →

Before and after

Today's inspection versus Edge vision at the line exit

Aspect	Visual inspection	With iLEAN Edge
Coverage	Depends on fatigue and cadence	Every container, every shift
Weeping at the neck	Seen if it drips in front of someone	Detected as moisture on the shoulder
Unreadable or crooked label	Caught at the branch	Rejected before wrapping
Where the defect surfaces	Inside a hazmat truck	On the reject conveyor
Evidence	None	Image linked to the batch
Borderline container	It ships	Escalated to a person

from defects caught at destination to defects caught on the line · from inspection dependent on human attention to continuous inspection · from in-transit spills to incidents avoided.

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 6-12 months, depending on packaging cadence.
• At least 30 % fewer defective-container claims and transport spill incidents, which is the figure to validate against last year's claims log.
• Avoided cost of trailer clean-up, replacement product and re-delivery, none of which appear in the packaging line's own budget, which is why the case is usually approved by logistics and safety before production.
• And a shipment that is no longer rejected in full because of one drum, which is where the customer relationship is actually measured.

at least 30% fewer defective-container claims and transport spill incidents, plus the avoided cost of clean-up and replacement. Estimated payback of 6 to 12 months depending on packaging cadence. *Estimate to validate.*

And the fair question from the production manager

“What if it rejects good containers?” — the false positive is the real risk of any vision system on a line, which is why the model is trained on good and defective containers of your own packaging, with the lighting of that line exit, not on a generic model. Classifying a cap or a label against a known reference is an anchored task where the best models drop below 1.5% error [1], and a borderline container is not simply rejected: it is escalated to the operator, whose decision feeds back into training.

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