The defect that defines your pellet is tiny: it has to be seen early

The problem

Hundreds of kilos can go by between one sample and the next.

At the pelletiser's real throughput, control is done by sampling: somebody takes a handful every so often and looks at it on a white tray. Hundreds of kilos can pass between samples, and the outcome depends on fatigue and on the light at that workstation. Whatever slips through goes into the big-bag, the big-bag onto the truck, and the truck to the customer.

• At the pelletiser's real rate, control is by sampling: somebody takes a handful now and then and looks at it on a white tray.
• Hundreds of kilos can go by between one sample and the next.
• The result depends on tiredness and on the light at that station.
• What gets through goes into the big bag, the big bag onto the truck and the truck to the customer.

How it fits the IRIS system

Edge on the belt, with local inference — and two outputs, not one.

Edge puts an overhead industrial camera above the belt at the cutter outlet, and a second one over the extruded profile. Inference runs locally, on a GPU at the line: no images to the cloud, because at this throughput the latency, the cost and the bandwidth would never allow it. The model is trained on good and bad samples of this plant's specific grade. And it has two outputs, not one. The first is reactive: it diverts out-of-spec material before it reaches the big-bag. The second is the one that actually changes the outcome: the trend curve of defect density. Because black specks appear progressively. Caught early, a purge or a filter change is enough. Caught at the end, the whole batch is gone. That is the conceptual jump: from reacting to pre-empting. The camera is not there to catch the defect, it is there to warn that the process is drifting while it can still be corrected with a five-minute purge instead of with a downgraded batch.

The first output is reactive: it diverts out-of-spec material before it reaches the big bag. The second is the one that changes the business: the defect rate becomes a process signal, not a final verdict.

See the full IRIS architecture →

Before and after

Pellet control, before and after

Aspect	Today	With iLEAN Edge
Coverage of the control	1-2% sampling	Continuous over the flow
Between one sample and the next	Hundreds of kilos unlooked at	Nothing goes by unlooked at
What the result depends on	Tiredness and station lighting	The model, the same at hour six
What happens to the defect	Caught at the end, if at all	Diverted before the big bag
The defect rate	A final verdict	A live process signal
Dependence on network or cloud	—	None: GPU at the line

Impact estimate

Impact estimate — to validate against 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, directly proportional to your current rework and grade-downgrade rate.
• From 1-2% sampling to continuous control of the flow.
• Less out-of-spec material reaching the big bag, which is where it stops being cheap to recover.
• And a continuous process signal that does not exist today: defect rate by batch and by shift.

Payback 5-12 months · from 1-2% sampling to continuous control of the flow. Estimated payback is five to twelve months and it depends directly on the current rate of rework and grade downgrading. It is a direct cut in yield loss, which in this business is margin, plus a cut in customer complaints. An estimate to validate against the kilos reprocessed per month today. Conservative ranges, an estimate to validate against the plant's real data.

And the fair question from the production manager

"How many bad samples does it need to be trained?" — fewer than feared, because the model is trained on good and bad samples of this plant's specific grade, not a generic catalogue. And it does not decide alone: it diverts on a threshold, and the person sets the threshold.

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