Butter and margarine control with AI — a badly churned mold goes unseen until it is already in the cold room.

The problem

Three variables decided at once and measured separately.

The quality of a stick of butter or a block of margarine depends on three things that intersect within the same window of minutes, yet are measured in different systems:

1. Churning — the churn (or the crystallization cylinder in margarine) decides the phase inversion. If the temperature climbs out of band or the residence time falls short, the whey droplets stay too large and the product will weep in the cold room or on the shelf.
2. Dosing — salt, vitamin A/D (margarine), lactic cultures, added water. A tenth of a point outside the range changes the sensory profile and, at the extreme, pushes you out of the legal category (butter ≥80% milk fat; three-quarter-fat 60-62%; light 39-41%, per EU Regulation 1308/2013).
3. Molding and packing — weight per portion, symmetry, color, integrity of the foil wrap or the tub. The operator checks them in passing — and in passing is exactly when the marginal units slip through.

The lab confirms the problem four hours later, when the batch is already in the cold room. The retailer's return confirms it a week later, when there is nothing left to do. The classic system works 99% of the time — the 1% that fails is what reaches the customer with whey outside the wrap.

How it fits the IRIS system

iLEAN does not add a fourth system — it seals the cracks between the churn, the ERP and packing.

The problem is not a lack of information: it is information living in islands that never gets cross-referenced at the critical moment. The churn speaks an old protocol. The ERP holds the recipe. The operator looks at the package. Each does its own job, but nobody crosses all three at once. iLEAN acts as the putty that fills that gap, without asking you to change the churn or the ERP.

Edge sees the package at the molding outlet. Connect reads the old churn and the recipe wherever it lives. The agent cross-references, anticipates the drift and holds the batch before the cold room. The person signs — never the other way around.

The three iLEAN pieces applied to butter and margarine control:

• Edge — a machine-vision (CNN) terminal over the molding and packing line. It reads package symmetry, color, the absence of weeping whey droplets, the integrity of the foil wrap or the tub closure, and compares against the pattern of the active SKU. If something does not add up it fires the actuator (ejector, warning light) in milliseconds. It works without a network: the line does not go blind even if the WiFi drops.
• Connect — captures the batch recipe whether it comes from the ERP, the vertical MES, or the Excel sheet where R&D updates vitamins A and D for margarine. It reads the old churn every few seconds (churning temperature, motor torque, residence time) without forcing you to replace it. And it captures what arrives from outside (a rennet change, an alert from the cream supplier, a new rule from the US retailer) at second zero.
• Agent — cross-references the recipe, the churn data, the line's history and the image of the package. If the drift points to weeping or weight out of band, it does not send an email at 10 pm: it holds the batch and alerts the quality manager through whatever channel they use. The person validates and signs; the line does not restart on its own.

See the full IRIS architecture →

Before and after

Classic butter/margarine control vs. cross-referenced with iLEAN

Aspect	Classic control (lab + eye)	With iLEAN Edge + Connect + Agent
Churn readings	Manual once per shift, noted on a log sheet	Continuous data, cross-referenced with the recipe
Weeping detection	The operator in passing, or the cold room hours later	Edge at the molding outlet, in milliseconds
Weight control per portion	Statistical sampling, hours of delay	100% of packages, with ejection if out of band
SKU changeover (250g → 125g, butter → margarine)	Paper checklist and the shift lead's signature	Recipe + visual template + dosing verified at once
Fat-content labeling	SKU label, no cross-check with actual dosing	Label verified against the batch's real recipe and churning
File for the IFS/BRC auditor	Rebuilt by hand, several weeks	Per-batch dossier, automatic, with a photo of the package

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 plant, 2 molding and packing lines, mix of 250g sticks and single portions, churn with an old PLC not integrated with the MES.
• Edge pilot on one line (camera over the molding outlet + actuator + integration with the batch recipe and the churn). First value expected within a few weeks, with an initial ability to hold packages that fall outside the pattern.
• Indicative payback between 4 and 9 months, depending on the frequency of weeping incidents, out-of-band weights and retailer complaints in recent years.
• The hard levers: scrap reduction (packages with weeping or underweight) ≥ 30%, a single recall avoided over a mismatched label, and fewer returns from the cold room.

And the quality director's reasonable doubt

"What if the AI holds good packages?" — the false-positive risk is real, which is why iLEAN is designed to assist and simplify, not to replace the manager's eye. In anchored tasks (reading the image of the package and comparing it with the recipe and the churn readings), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds the batch and the person signs. The three safety rings exist precisely for this.

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

How it connects with other solutions in the dairy and fats cluster: milk fat control · baby food glass jar control · yogurt fermentation control.