The camera rejects the bad seal before packing

The problem

The human eye cannot hold up at the line's real pace.

100% human visual control is impossible at real line speed, more so with several filling formulas rotating through the shift, each behaving differently when sealed. Sampling-based inspection leaves gaps that end in complaints or a full tray return from the customer.

• A badly closed crimp is not always visible: an empanada with the edge barely open passes visual control, bursts in the tunnel oven or opens during freezing and stains the rest of the tray.
• 100% human visual control is impossible at real line speed, and fatigue does the rest from the third hour of the shift: the judgment applied to piece number one thousand is not the one applied to piece number fifty.
• With several filling formulas rotating in the same shift, each behaves differently when sealed: the wetter filling escapes through the edge and the colder one keeps the dough from bonding. What the operator learned to look for with one recipe stops serving them with the next.
• Sampling-based inspection leaves gaps that end in a complaint or a full tray return from the customer, which is where the cost multiplies: you do not lose one empanada, you lose the pack, the film, the process and the trust of the chain that bought it.

How it fits the IRIS system

Edge — piece-by-piece classification in milliseconds, before packing.

Edge: local CNN vision over the forming-sealing line, no cloud dependency. The model, trained on real examples of each formula, infers in milliseconds and rejects the defective piece before packing.

Inference is local, on the line itself, with no dependency on the plant network or the cloud. At the pace of a forming machine that is not a technical detail: it is the difference between rejecting the bad piece in the next meter of conveyor and finding it at the customer, two weeks later, with a whole tray compromised.

See the full IRIS architecture →

Before and after

Today's inspection versus Edge vision

Aspect	Visual inspection	With iLEAN Edge
Coverage	Sampling and fatigue	100% of pieces, constant
A barely open crimp	Slips through	Classified piece by piece
Point of detection	The customer or final inspection	The conveyor, before packing
Several formulas in a shift	Each one fools you differently	One model per active formula
Tray stained by one piece	Lost whole	Never gets formed
Edge cases	Decided by haste	Escalated to a person

Partial sampling → 100% piece inspection. Sealing-defect complaints → drastic reduction.

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 5-12 months, depending on your current sealing scrap ratio.
• From partial sampling to 100% inspection of the pieces leaving the forming machine.
• A drastic reduction in complaints and returns for sealing defects, which is what the retail chain actually measures you on.
• And the tray that used to be lost whole because one single piece opened during freezing stops reaching packing.

Estimated payback 5-12 months depending on current scrap ratio. *Estimate to validate*.

And the fair question from the production manager

“What if it rejects good empanadas?” — the false positive is the real risk of any vision system, which is why the model is trained on good and bad pieces from your specific formulas and with the lighting of that station, not on a generic model. Edge cases are not simply rejected: they are escalated for a person to decide, and every decision feeds back into training. Classification is an anchored task, where the best models drop below 1.5% error [1].

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