The pasteurizer stays the same — but now it talks

The problem

The data food safety depends on is the data that circulates worst.

Did this milk reach 72 °C? Was there a flow diversion? How many times and for how long? Those are the questions defining the food safety of everything leaving the plant, and today they are answered on a paper roll:

• Proving a batch's curve forces reconstructing the thermal archive — by hand, finding the chart stretch matching that hour and cross-referencing it with what was passing through. In an audit, that is hours.
• Correlating a diversion with a microbiological deviation is archaeology — the diversion was marked on the paper, but knowing which milk was passing at that instant demands a manual exercise that rarely adds up one hundred percent.
• The obvious alternative is not viable — replacing the pasteurizer is hundreds of thousands of euros, weeks of stoppage and revalidating the equipment. And touching its program is not either: nobody wants to assume the regulatory risk of intervening in a validated machine.

So the plant lives with its most critical data on a medium that cannot be queried, and every audit and every investigation starts with the same manual work from scratch.

How it fits the IRIS system

Connect in photo mode on the panel — looking at the screen, not intervening in the machine.

The missing information is already on the screen: the panel shows it every cycle. What does not exist is the path from that glass to the central system. Connect solves it from the side that touches nothing — by looking — and therefore without opening the equipment's network, without modifying its program and without needing to revalidate it.

A fixed industrial camera aimed at the screen, rated for a wet area. Non-intrusive installation: nothing is connected to the machine. Optical recognition with an anchored model builds the time series of outlet temperature, flow rate and flow diversion events, and cross-references it with the ERP and the vat sheet to know which milk was passing at each moment.

How Connect operates on the pasteurizer's panel:

• Non-intrusive installation, rated for a wet area — a mount is installed in front of the screen, with protection suited to the environment. Nothing is connected to the equipment: not the bus, not the cabinet, not the program. The pasteurizer's validation stays intact.
• OCR anchored to that specific panel — the model knows that screen's real layout: where the outlet temperature is, where the flow rate, where the differential pressure and where the diversion indicator. It reads values at known positions.
• A time series, not snapshots — the continuous reading is reconstructed as the cycle's curve. That is what lets you see how long a diversion lasted, not just that it happened.
• Associated with the batch, not the clock — the cross-reference with the ERP and the vat sheet says which milk was passing at each moment, so the curve stays tied to the batch. Without that cross-reference the series is a chart; with it, it is traceability.
• It feeds the auditable evidence pack — the batch's curve enters the dossier that later gets downloaded, instead of being reconstructed from the paper archive every time someone asks.

See the full IRIS architecture →

Before and after

Paper thermal record vs. curve digitized with Connect

Aspect	Archived paper chart	With iLEAN Connect on the panel
The batch's thermal curve	On paper, archived	A time series queryable by batch
Proving pasteurization to an auditor	Hours of archive work	Seconds
Detection of an anomalous flow diversion	"When reviewing the chart"	In the moment
Diversion ↔ microbiological deviation correlation	Documentary archaeology	A cross-referenced query per batch
Intervention on the equipment	Any integration demands touching and revalidating it	None — external camera, validation intact
Pasteurizer replacement	Hundreds of thousands of euros and weeks of stoppage	Deferrable indefinitely

Impact estimate

Impact estimate for your plant — to be validated 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.

• Dairy with a 12-to-20-year-old HTST pasteurizer, a network-isolated panel and a thermal record on a paper chart.
• Connect pilot on the panel: a wet-area-rated industrial camera, model anchoring to that specific screen and cross-reference with the ERP and the vat sheet. First value expected within a few weeks.
• Indicative payback between 5 and 10 months. Estimate to be validated.
• Add the indefinite deferral of the pasteurizer replacement CAPEX, which is usually the figure that moves the committee: the project is postponed without giving up the data that justified it.
• And the reduction of the risk of a major nonconformity for incomplete thermal traceability, which in a dairy is not a documentary matter: it is the permission to sell to certain customers.

And the fair question from the quality manager

"Does this force me to revalidate the pasteurizer?" — no. The camera is an independent device looking at the screen from outside: it does not connect to the equipment, shares no wiring and writes nothing into its logic. For the pasteurizer nothing has changed, so its validation remains in force. On the reading, hallucination is a problem of free generation, not of anchored tasks: reading a value at a known position on a known panel is where the best models brought the error below 1.5% ^[1], with time-redundant reading and discarding of frames that do not reach the confidence threshold.

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