Yogurt line control with iLEAN Edge — 600 cups/min, every one inspected, none out of spec through to palletizing.

The problem

High speed, multi-lane, one-off defect: human inspection cannot keep up.

A standard yogurt line runs multi-lane at several hundred cups per minute. The operator at the end of the line cannot see every cup — it is physically impossible. What they do, and do well, is check random samples and keep an eye on the SCADA panel. The defects that get through are the ones that slip between samples:

• Fill level out of spec — the "underfilled" cup that reaches the customer and triggers a complaint. Under the retailer's rules, there is a tolerance threshold per unit and per batch.
• Defective sealing — a fold in the top film, a trace of yogurt between lid and cup, a weld gone wrong because of temperature. Visible to a camera, nearly invisible to the eye at 600/min.
• Microleak — invisible to the eye and to the camera alone. The product reaches the shelf and three days later the tray smells bad — mass returns, brand damage.

The classic system (samples + checklist + complaint) works 99% of the time — and that 1% is the batch pulled from the shelf. As in other industries, this is not about getting rid of the operator; it is about freeing them from dumb work (looking at what was fine) so they can do the work that matters (deciding on what does not add up).

How it fits the IRIS system

iLEAN does not add another layer between the filler and the rejector — it seals the crack between them.

Your yogurt line already has a filler, a heat sealer, a rejector and a SCADA. The problem is not a lack of machines; it is that each one does its own job and nobody cross-references the results in real time. iLEAN acts as the putty that fills the cracks between those pieces, reads every cup, decides in milliseconds and fires the rejector you already have — replacing nothing.

Edge sees every cup at 600/min. The agent cross-references the batch and the SCADA. When a defect pattern appears, it proposes the root cause to the shift lead. The person signs — the line does not restart on its own.

The three iLEAN pieces applied to the yogurt line:

• Edge — terminals with machine vision (CNN) mounted over the filler and the heat sealer, integrated with the vacuum test chamber. It inspects fill level, sealing, leaks, coding. It fires whatever actuator you already have, in milliseconds. It works without a network: if the plant loses WiFi, the camera keeps inspecting and rejecting on the power from the cabinet.
• Connect — captures the data that arrives from outside the line (a lid-batch change from the supplier that purchasing flagged by email, a recipe adjustment from the R&D lead in an Excel file, new retailer parameters over WhatsApp) at second zero, with no forwarding and no meetings.
• Agent — cross-references the rejection rate by lane, by hour, by batch, against the heat sealer parameters and the recipe. If the rate climbs on line 3 at 2:20 pm, it does not send an email at 10 pm: it alerts the shift lead with the probable cause. The person decides and signs — Agents propose, the person executes.

See the full IRIS architecture →

Before and after

Sample-based inspection vs. continuous AI vision inspection

Aspect	Operator + samples + SCADA	With iLEAN Edge + Agent
Coverage	Random samples (typically <1%)	Every cup, every lane
Defect detection	Reactive: shows up in sampling or in a complaint	At second zero, before palletizing
Rejection	Suspect batch put on hold	Cup by cup, without stopping the line
Root cause	Reconstructed after the fact	Agent cross-references the SCADA and proposes it
Per-batch dossier	Quality sheet filled in at the end	Automatic, with a photo of the defect
Operation without a network	n/a	Edge keeps rejecting on the power from the cabinet

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Multi-lane packing line at several hundred cups/min, standard heat sealer, existing pneumatic rejector. Edge pilot on one line.
• First value expected within a few weeks: fill level + sealing operational with a per-batch dossier; leaks depending on whether a vacuum test chamber already exists or is added.
• Indicative payback between 4 and 9 months, depending on the current rate of defects detected late, the average cost of a retailer return and the brand cost of a microleak reaching the shelf.
• Hard levers: reduction of defects reaching the customer ≥ 30% (conservative estimate); a single avoided shelf recall pays for the pilot; operator time freed for tasks that genuinely require judgment.

And the operations director's reasonable doubt

"What if the camera drifts out of adjustment and starts rejecting good cups?" — positive preemption: the agent monitors the system's own false-positive rate and raises a flag before it gets out of hand. The quality manager sees why each cup is rejected, adjusts the sensitivity without touching code, and the system learns from the corrections. In anchored tasks (looking at a cup and comparing it with a pattern), the best models brought error below 1.5% ^[2]. And even so, what is critical (stopping the line, disqualifying a whole batch) is never decided alone: the agent proposes, the person signs.

^[1] Automotive quality standard on the order of 25 PPM — Symestic.

^[2] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.