Coffee capsule sealing defects — the customer discovers them in their kitchen, and returns the lot.

The problem

The capsule nobody sees is the capsule the retailer returns.

A single-serve compatible capsule line is a high-speed, low-tolerance system: the rim geometry, the heat sealer's temperature, the closing pressure and the coffee weight have to line up to the millimetre. And meanwhile:

1. The heat sealer ages at its own pace. A thermocouple that drifts two tenths for a quarter of an hour goes unnoticed in SCADA — and leaves thousands of capsules with sealing in debt that the naked eye cannot spot.
2. Quality control is statistical. One in every 5,000 is opened and measured. If the problem started 4,999 capsules ago, they are all already packed, all on the pallet, all on their way to the distributor.
3. The consumer is the final quality control. The capsule does not brew well, the coffee comes out watery, the online review is one star, and the retailer asks for a lot recall. The brand damage weighs more than the product cost.

The quality manager knows this, but at 1,200 capsules/min no human eye can keep up. The classic system (thermocouple + statistical control + checklist) works 99% of the time. That 1% is the retailer's return — and each time it is hard product cost, hard reverse-logistics cost, brand cost, and a crisis committee.

How it fits into the IRIS system

iLEAN does not add a fourth system — it seals the crack between the heat sealer, the SCADA and the packing.

The problem with faulty sealing is not a lack of technology on the line: it is information living in islands (the heat sealer has its panel, the SCADA has its history, packing has its counter) that, at the critical moment — the defective capsule crossing towards packing — does not reach the actuator in time. iLEAN is the filler that closes that gap without asking you to change the heat sealer or the cartoner.

Vision sees the capsule before packing. Connect reads the heat sealer even if it is old. The agent cross-checks visual quality with process conditions and proposes the adjustment. The person signs off — never the other way round.

The three iLEAN pieces applied to coffee capsule sealing:

• iLEAN Vision (Edge) — a terminal with machine vision (CNN) over the heat sealer's exit. It sees every capsule at 1,200/min, detects incomplete sealing, aluminium deformation, a foreign particle on the rim, defective printing. It triggers the ejector (air blast or pneumatic arm) in ~45 ms. It works with no network: if the plant loses its connection, Edge keeps seeing and rejecting.
• iLEAN Connect — reads the heat sealer even if it is an old machine with an isolated local computer or an analogue panel (a photo every few minutes solves the capture). And it captures what comes from outside: the retailer's email with the first return, the salesperson's WhatsApp warning about a social-media thread, all at second zero, without anyone forwarding anything.
• iLEAN Agents — cross-checks the Edge's visual data (which type of defect) with the process conditions (temperature, pressure, pace) Connect has read from the heat sealer, and reasons the root cause. When it detects a trend, it does not send an email: it proposes the adjustment to the line leader with an argument ("the temperature has dropped two tenths in the last 12 minutes, we recommend checking thermocouple 3 before raising the pace"). The person signs off — the line never readjusts itself.

See the full IRIS architecture →

Before and after

Statistical control vs. 100% inspection with iLEAN Vision

Aspect	Thermocouple + statistical control	With iLEAN Vision + Connect + Agent
Inspection coverage	1 in every 5,000 capsules	100% of capsules at 1,200/min
Defect detection	When the sample is opened (minutes later)	In line, before packing, in milliseconds
Root cause	Later investigation, with no image of the defect	Defect photo cross-checked with process conditions
Recall from a retailer return	Present risk	Risk pushed to the residual
Operation with no network	n/a	Edge keeps rejecting on cabinet power
File for the BRC / IFS auditor	Reconstructed by hand, with samples	Dossier per lot with a photo of every rejected capsule

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 figures of your plant. We set it out so the committee has an order of magnitude; we refine it during the diagnostic.

• Single-serve compatible line at 1,000-1,500 capsules/min, heat sealer with temperature control, cartoner downstream. A documented history of sealing returns from the last year.
• Edge pilot on one line with a camera at the heat sealer's exit + pneumatic ejector. First expected value within a few weeks: detection above 90%, rejection before packing.
• A reasonable reduction in defective capsules packed of ≥ 30% in the first quarter, with a clear trend towards the line's limit as the model trained on your own images improves.
• Indicative payback between 4 and 9 months, depending on the cost of avoided returns and the frequency of retailer incidents. A single avoided lot recall pays for the pilot.

And the quality director's reasonable doubt

"What if the AI discards a good capsule by mistake and fills my pallet with pointless rejects?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI simply classifies an image against a reference (well-sealed vs. badly sealed capsule), the best models brought the error below 1.5% ^[1]. And automotive quality demands a standard of 25 PPM ^[2] — compatible coffee for the end consumer asks for something similar. iLEAN trains on your images and improves; the person validates the learning curve. The three safety rings are there for exactly this.

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

^[2] Symestic — automotive quality standard in the order of 25 PPM (parts per million).