Sugar refinery with AI — the commercial grade is not decided in the lab, it is decided in the vacuum pan.

The problem

The pan has already boiled and nobody told the operator the syrup had changed.

Refined sugar quality is decided upstream — in the vacuum pan, when the operator feeds the syrup, introduces the seed and lets the grain grow. If that strike drifts, what comes out of the centrifuge cannot be recovered: it is either reworked or downgraded to a lower commercial grade. And most of the time, the operator does not know it is drifting until the LIMS, an hour later, returns the polarimeter result.

It happens because the information arrives late and through different channels:

1. The color of the mother liquor charged — depends on how the earlier stages evaporated; the data lives on the evaporator console and nobody carries it to the pan panel.
2. The veteran's adjustments — the pan's sight glass, the feel of the syrup flow, the moment of crystallization. That eye is not in any system.
3. The lab result — polarization, ICUMSA color, ash. It arrives fine — but an hour later, when that pan has already discharged and is on its way to packing.

Everyone does their job. The system fails in the coordination, not in the measurements — and the loss is real: a season with 3-4% of batches downgraded to a lower commercial grade is hundreds of thousands of euros in refined sugar sold as standard white.

How it fits the IRIS system

iLEAN does not replace the LIMS or the pan operator — it seals the cracks between what you already have.

The problem with refinery control is not a lack of information: it is information living in islands that at the critical moment (the next pan starting up before the lab returns the previous one) does not reach the decision-maker in time. iLEAN acts as the putty that fills those gaps, without asking you to change the pan console, the LIMS or the dryer.

Edge watches color and grain at the dryer outlet. Connect reads the pan console and the LIMS result at second zero. The agent correlates and proposes the adjustment to the operator. The person signs — never the other way around.

The three iLEAN pieces applied to refined sugar quality control:

• Edge — a terminal with machine vision (CNN) over the sugar belt coming out of the dryer or the holding silo. It reads color, mean grain size and grain-size dispersion without touching the line. If the reading drifts from the commercial grade declared for that batch, it fires an actuator (signal light, diverter to the standard-white silo) in milliseconds. It works without a network: if the plant loses its WiFi, Edge keeps sorting.
• Connect — captures the pan data whether it comes from a modern console or from an old PLC with only a local screen. It reads the LIMS as soon as the lab issues the result, without waiting for the shift report. And it also captures what arrives from outside (a retailer spec changed in an email, a complaint from an industrial customer over caking) and puts it in the system at second zero.
• Agent — cross-checks the Edge reading at the outlet, the console of the pan that produced that batch, the LIMS result and the season history. If it detects that the next strike is going to come out with a deviation, it does not send an email: it alerts the operator through whatever channel they use, with the specific proposed adjustment (raise the vacuum 3 mbar, lower the syrup flow, delay the seed). The operator validates and signs; the pan does not adjust itself.

See the full IRIS architecture →

Before and after

Classic control vs. cross-checked control with iLEAN

Aspect	Classic control (LIMS + veteran operator)	With iLEAN Edge + Connect + Agent
ICUMSA color	Lab result every hour, post-centrifuge	In-line reading + correlation with the pan that produced the batch
Polarization	Polarimeter reading, logged in an isolated LIMS	Cross-checked with the pan console as soon as it is issued
Final moisture	Sample at packing, too late to react	Anticipated from the dryer reading + proposed adjustment
Adjusting the next pan	Operator's memory + sight-glass intuition	Concrete proposal from the agent, validated by the operator
Commercial-grade downgrade	Confirmed at packing, no reaction possible	Diverted to the standard-white silo before packing
Season file	Rebuilt by hand at every season close	Per-pan dossier, automatic, with photos and traces

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.

• Mid-sized refinery with 2-4 vacuum pans, a single dryer outlet, a mix of premium refined + standard white + industrial.
• Edge pilot at the dryer outlet (color + grain camera + integration with the pan console and LIMS). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the historical percentage of batches downgraded in the previous season and the €/kg spread between your premium refined and standard white.
• Hard lever: ≥ 30% reduction in batches downgraded to a lower commercial grade. Over a season of several million kilos, a single well-captured correlation pays for the pilot.

And the refinery lead's reasonable doubt

"What if the agent proposes the wrong adjustment and I ruin the pan?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely correlates what is already measured (color, vacuum, polarization, feed profile), the best models brought error below 1.5% ^[1]. And even so, the pan adjustment is not decided by the AI: the agent proposes, the operator signs. The three safety rings exist precisely for this.

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

Flour mill control with AI → · Sunflower oil refining with AI →