Flour milling with AI — a low falling number is not an incident, it is a complaint from the baker.

The problem

The wheat changes with every truck and the blend recipe changes in a spreadsheet.

The quality of an industrial flour depends on a balance that breaks every time a new wheat delivery arrives: different harvest moisture, different hardness, different protein profile. Conditioning (tempering) should compensate for it, but it only does so properly if the incoming data arrives in time and is cross-referenced with the destination of the blend. And that rarely happens.

It happens because the information lives in four different places:

1. Moisture and hardness of the wheat on the truck — on the delivery note, in the intake lab, sometimes in the purchasing manager's head.
2. Milling stand parameters — in the mill's SCADA, a screen the shift lead looks at every couple of hours.
3. Particle size and color at the outlet — the plansifter does its job, but the fine analysis arrives later.
4. Falling number, protein, ash — in the in-house lab, an hour later, once the flour is already in the silo.

When the industrial baker calls at 11 the next morning to say the dough is sticking, that batch is already spread across 30 points of sale. The mill works 95% of the time. That 5% is what comes back as a complaint — and every complaint is hard cost plus the cost to the account.

How it fits the IRIS system

iLEAN does not add a fifth system — it seals the cracks between the four you already have.

The problem with blend control in a mill is not a lack of information: it is information living on islands that, at the critical moment (the new truck discharging while the day's blend keeps running to the silo), does not reach the decision-maker in time. iLEAN acts as the putty that fills those gaps, without asking you to change the mill's SCADA, the LIMS or the silo.

Edge sees the flour at the plansifter outlet. Connect reads the wheat delivery note, the SCADA and the LIMS at second zero. The agent correlates and proposes the adjustment to the shift lead. The person signs — never the other way round.

The three iLEAN pieces applied to quality control in a flour mill:

• Edge — a terminal with machine vision (CNN) over the plansifter outlet belt or the loading silo. It reads color (correlated with ash), mean particle size and dispersion without touching the line. If the reading drifts away from the declared type (T55 coming out darker than the range, coarser particle size), it triggers a diversion to another silo or requests a conditioning adjustment. It works without a network: Edge keeps classifying even if the plant loses WiFi.
• Connect — captures the incoming wheat delivery note whether it comes from the ERP, from the intake lab or from a spreadsheet kept by the purchasing manager. It reads the mill SCADA and the LIMS as soon as they issue a result, without waiting for the shift report. And it also captures what arrives from outside (a spec change from the industrial baker by email, a customer complaint over WhatsApp) at second zero.
• Agent — cross-references the incoming wheat, the conditioning parameters, the Edge reading at the outlet, the LIMS result and the blend recipe of the destination silo. If it detects that the batch in progress is going to come out off type, it does not send an email at 10 p.m.: it alerts the shift lead on whichever channel they use, with a concrete proposal (raise conditioning moisture, change the stand blend, redirect to a silo of a different type). The lead validates and signs; the blend is never adjusted on its own.

See the full IRIS architecture →

Before and after

Classic control vs. cross-referenced control with iLEAN

Aspect	Classic control (LIMS + shift lead)	With iLEAN Edge + Connect + Agent
New wheat intake	Delivery note to the lab, result hours later	Captured at second zero + cross-referenced with the SCADA
Particle size / color	Spot sampling + manual sieving	Continuous inline reading, on every batch
Conditioning adjustment	The miller's memory, based on the last analysis	Agent proposal cross-referencing wheat + SCADA + Edge
Type deviation (T55 → T80)	Confirmed by the lab, already in the silo	Diverted to the right silo before discharge
Baker complaint	Investigation after the fact, days to rebuild the batch	Per-batch dossier with every parameter, automatic
Shift close-out	The shift lead's manual spreadsheet	Automatic report with deviations and adjustments

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 mill. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Mid-sized mill with 3-4 milling stands, a blend of soft and hard wheat, flours for industrial bakery + biscuits + pastry.
• Edge pilot at the main plansifter outlet (color + particle size camera + integration with SCADA and LIMS). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the frequency of complaints for falling number / off-type ash recorded over the last few years and the average weight of the batch claimed.
• Hard lever: a reduction of ≥ 30% in off-type batches redirected to secondary uses. One single major complaint from an industrial baker pays for the pilot.

And the mill manager's reasonable doubt

“What if the agent proposes the wrong blend and I ruin a whole silo?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely correlates what is already measured (wheat moisture, stand parameters, the plansifter reading, falling number), the best models brought error below 1.5% ^[1]. And even so, the silo blend is not decided by the AI: the agent proposes, the miller signs. The three safety rings exist precisely for this.

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

Sugar refinery control with AI → · Rice whitening control with AI →