Fixed-weight cheese slicing — bring giveaway down without risking underweight.

The problem

The cheese block is not uniform — and the slicer acts as if it were.

A portion of sliced cheese carries a declared weight on the pack — 100 g, 150 g, 200 g. European rules on nominal quantities require that no portion systematically falls below the declared weight. The plant protects itself against underweight by nudging the setting upwards: cutting slices a little thicker than strictly necessary to reach the target. Across the plant as a whole, that means giving grams away in every portion just to be sure of compliance.

The operational problem: a block of cured cheese is not uniform. Three known causes:

1. Variable density — ripening leaves some areas more compact and others more porous within the same block.
2. Warping and geometry — the block comes out of the cold room with a curve; it is not a perfect rectangular solid.
3. Internal eyes and rind — a block with an internal eye weighs less per slice than a compact one in the same area.

The slicer cuts at a fixed thickness, assuming the block is uniform. The multihead weigher or the tray sealer downstream corrects after the fact by adding slices. When the result goes over the declared weight, the system works — but it gives product away. Across millions of portions a year, every gram given away is hundreds of thousands of euros of margin.

How it fits the IRIS system

iLEAN Vision 3D does not replace the slicer — it hands it the data on the block, slice by slice.

The giveaway problem is not the slicer — the slicer does what it is told. The problem is that nobody is telling it what the block in front of it looks like at any given moment. iLEAN acts as the putty between the real geometry of the block, the slicer's settings and the output to the weigher/tray sealer — without asking you to change the slicer.

The scanner measures the block. The agent calculates thickness slice by slice. The slicer cuts exactly what is needed. The portion meets its weight with no underweight and no giveaway. The operator is still in charge.

The iLEAN pieces applied to fixed-weight slicing:

• Edge — 3D scanner over the incoming block. Ahead of the blade, a laser or depth scanner measures the geometry of the block (length, warping, usable cross-section). A CNN estimates density by area from the SKU's history plus the visible rind. The data reaches the agent in milliseconds.
• Connect — capture of the real weight per portion from the multihead weigher or the tray sealer downstream. Real-time feedback to the agent: real average weight, spread, percentage of portions below and above the declared value.
• Agent. Cross-references block geometry + estimated density + the declared fixed portion weight and calculates the thickness set-point for each slice. It passes that to the slicer via retrofit (if it does not natively expose a dynamic set-point) or via the native interface if it does. It learns from the real weight measured downstream and adjusts the density model per SKU. The agent does not approve the portion: the weigher verifies it. The three rings: the agent proposes the set-point, the physical weigher validates it, the operator has the stop button.

See the full IRIS architecture →

Before and after

Fixed-thickness slicing vs. adaptive cutting with iLEAN Vision 3D

Aspect	Fixed thickness + downstream correction	With iLEAN Vision 3D (Edge + agent)
Assumption about the block	Uniform	Geometry and density measured, slice by slice
Average giveaway	A wide margin to avoid underweight	Minimal margin — the slice adapts to the real block
Underweight risk	Controlled by overweight	Controlled by multihead weigher verification
Block warping	Causes weight spread	The scanner measures it and the agent compensates
SKU changeover	Manual reprogramming	The agent applies the SKU's density model
Operation without a network	n/a	Edge keeps measuring and setting the set-point on the panel's power

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

• Cured or semi-cured cheese slicing line, fixed-weight portions (100 g / 150 g / 200 g), multihead weigher and tray sealer downstream.
• iLEAN Vision 3D pilot on the slicer: 3D scanner + set-point retrofit (where applicable) + integration with the weigher. First value within a few weeks: giveaway measured before and after.
• Indicative payback between 4 and 9 months, faster the higher the annual volume and the higher the starting giveaway.
• Expected reduction in giveaway per portion of ≥ 30% against the baseline, with no increase in underweight risk (the multihead weigher keeps verifying).
• The hard lever is the grams given away per portion × annual volume. On medium-to-high volume lines, that is hundreds of thousands of euros a year in recovered margin.

And the plant manager's reasonable doubt

“What if the agent gets the thickness wrong and we end up underweight?” — hallucination is a problem of free generation, not of anchored tasks. Turning a 3D geometry plus an estimated density into a thickness set-point is an anchored task: the best models brought error below 1.5% ^[1]. And even so, the multihead weigher physically verifies every portion downstream — the agent proposes, the weigher validates, the operator stops the line if needed. The three safety rings, applied to weight.

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