The plant's most expensive asset is also its most silent

The problem

The machine that decides IQF quality is the one that shares nothing.

The IQF quality of a breaded product depends on a curve nobody logs continuously today: that the pieces come out loose rather than clumped, that the core reaches temperature, that the coating does not crack. What exists today is two or three manual readings per shift on a form. Enough to comply, not enough to understand. When a distributor returns a pallet for clumped product, there is no way to know what the tunnel did that night. And the obvious way out — replacing the equipment with a connected one — does not hold up: nobody replaces a blast freezing tunnel in order to obtain a data point.

• The blast or spiral tunnel decides whether the croquettes come out loose or clumped, whether the core reaches temperature, whether the coating cracks. It decides all of that and logs none of it: proprietary controls, no network, ten to twenty-five years old.
• What exists today is two or three manual readings per shift on a temperature form. Enough to satisfy the record, not enough to understand a bad night.
• When a retail distributor returns a pallet for clumped product, nobody can say what the tunnel did during that batch: the curve was on the panel for a few seconds and then it was gone.
• And the obvious way out does not hold up: nobody replaces a freezing tunnel worth hundreds of thousands, with weeks of downtime, to obtain a data point.

How it fits the IRIS system

Connect in machine-panel photo mode — solved from outside, without opening the cabinet.

Connect in machine-panel photo mode.

It does not integrate with the tunnel's controller, opens no ports and installs nothing on the machine. On an asset this old and this critical, that is the difference between a case maintenance approves and a project nobody dares to authorize. And because every reading is tied to the work order that was inside the tunnel, the curve is not a machine log: it is a property of the batch, alongside its recipe and its lots.

• A fixed industrial camera with controlled lighting aimed at the panel.
• A vision model with grounded optical reading interprets every reading and builds the time series.
• The system cross-references the series with the active work order, so every stretch of the curve is tied to the batch that was inside the tunnel.
• The curve becomes available for root cause analysis and as an automatic audit record. The cabinet is never opened, the controller is never touched and the manufacturer's warranty is never voided.

See the full IRIS architecture →

Before and after

Today's tunnel versus the tunnel being read

Aspect	Today	With iLEAN Connect
The freezing curve	Two or three readings per shift	A continuous series per batch
A pallet returned for clumped product	"We do not know what happened"	Curve and complaint cross-referenced in minutes
Temperature record for the audit	A paper form	Automatic evidence tied to the batch
The tunnel's controller	—	Untouched, warranty intact
Replacement CAPEX	In the plan	Deferred indefinitely
Drift in the freezing process	Invisible until a complaint	Visible as a trend

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 5-10 months.
• Indefinite deferral of the tunnel replacement CAPEX, which is the largest single investment line in a precooked frozen food plant.
• Fewer returns for freezing quality, because a curve that drifted is seen before the pallet leaves rather than after the distributor complains.
• And the temperature record for the audit goes from a form somebody fills in to an automatic series nobody has to defend.

estimated payback of 5 to 10 months, plus indefinite deferral of the tunnel replacement CAPEX and a reduction in returns for freezing quality. *Estimate to be validated*.

And the fair question from the production manager

“Does a camera reading a panel hold up in front of an auditor?” — here the task is anchored: a fixed screen, digits in a known position and a physically possible range, where the best models drop below 1.5% error [1]. The system discards any frame whose value falls outside the possible range instead of accepting it, so a glare or a hand in front of the panel never becomes a data point. And every reading keeps its original frame as evidence.

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