Catch the defect before it reaches the customer

The problem

At line speed, the human eye loses to the borderline case.

an off-range color, a visible foreign body or a bad seal can reach the customer, triggering a complaint.

• An off-range color, a foreign body visible in a transparent pouch or a seal with powder trapped in it are the three defects that generate complaints in this sub-sector.
• All three are visible. The problem is that they are visible to somebody checking thousands of packs per shift, and fatigue lets exactly the borderline ones through.
• Color is the worst of the three, because it drifts slowly: the run starts inside the ASTA window and ends outside it, and nobody sees the change happen pack by pack. It is only obvious when two pouches from different ends of the run are put side by side, which happens at the customer.
• A trapped seal is not a cosmetic issue either: it is the entry route for moisture, and moisture in a spice pack means caking in the jar and a shelf life that does not reach its printed date.

How it fits the IRIS system

Edge — every pack classified in milliseconds, before sealing.

Edge camera + CNN trained on the specific product, millisecond inference, automatic rejection.

Catching it on the line instead of at the customer changes what the defect costs: a rejected pouch is grams of product, while a complaint is a pallet returned, an investigation opened, a corrective action to write and a commercial conversation you did not want to have. The unit cost of the same defect differs by three orders of magnitude depending on where it is found.

See the full IRIS architecture →

Before and after

Today's inspection versus Edge vision

Aspect	Visual sampling	With iLEAN Edge
Coverage	A sample, and it depends on fatigue	Every pack, at cadence
Slow color drift	Nobody sees it happen	Detected as a trend mid-run
Foreign body in the pouch	Caught if someone is looking	Classified pack by pack
Powder trapped in the seal	Found by the customer	Rejected before sealing
Point of detection	The complaint	The line itself
Borderline cases	They slip through	Escalated to a person

defects caught downstream → caught and rejected on the line itself.

Impact estimate

Estimated impact — to validate with your own 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 4-9 months, depending on your packing volume and your current complaint rate.
• Estimated reduction of at least 30 % in complaints and returns on the covered formats.
• Color drift caught mid-run instead of at the end, which is the difference between adjusting and reprocessing.
• And the sealing defect that shortens shelf life stops leaving the plant inside the pallet.

≥30% reduction in complaints and returns, estimated payback of 4-9 months. *Estimate to validate*.

And the fair question from the production manager

"What if it rejects good product?" — the false positive is the real risk of any vision system, which is why the model is trained on good and bad packs of your own formats and with that line's lighting, not on a generic reference. Classifying a pack against a known reference is an anchored task, where the best models drop below 1.5% error [1]. And borderline packs are not simply rejected: they are escalated for a person to decide, and every decision feeds back into training.

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