Meat cutting traceability — holding the carcass → cut link is the foundation of the whole business.

The problem

The carcass goes in whole and comes out as thirty cuts — and the link is held together by one operator and a shout.

A well-run cutting room is a small miracle of coordination. The carcass arrives hanging from the overhead rail, the operator lowers it onto the table, breakdown begins. Out of it come leg, shoulder, loin, rib rack, belly, trim… each to its own belt or cart, each with its own commercial destination. The problem: none of that is measured continuously.

1. Identifying the carcass at the start of breakdown is done by scanning the hook or by the operator calling it out — and sometimes, at peak hours, the scan is missed and it gets assumed.
2. Classifying each cut on the way out is done by the operator's eye — a veteran nails it; a seasonal temp, less so. And nobody records whether the cut was a leg or a shoulder, beyond what the weigh-labeler registers afterwards.
3. The carcass → cut link is held in someone's head, in the flow, and translated into the label at the end. If things get mixed up along the way, the link breaks — and that does not surface until a food safety alert tied to the slaughter batch arrives and specific cuts have to be recalled.

The meat quality manager knows the risk. But asking the operator to scan every single cut means breaking the room's rhythm — and rhythm is cost. So it rests on the veteran's eye. Which leaves the day they retire.

How it fits the IRIS system

iLEAN does not break the room's rhythm — it seals the crack between the carcass and the label on the cut.

The problem is not a lack of technology in the room: there is a weigh-labeler, there is an ERP, there is hook scanning. The problem is that those systems do not talk to each other continuously, and the carcass → cut link ends up hanging from the operator's memory. iLEAN acts as the putty that seals that crack, without asking you to change the scale, the ERP or your room's rhythm.

Vision recognizes every cut on the way out. Tracer holds the link to the carcass. The agent cross-checks with the scale and triggers the correct label. The operator carries on working exactly as before — the system only interrupts them if something does not add up.

The iLEAN pieces applied to meat cutting:

• iLEAN Vision (Edge with vision) — a terminal with an industrial camera above the outfeed belt or the sorting table. The convolutional neural network recognizes each cut in real time (leg, shoulder, loin, rib rack, belly…) without the operator doing anything differently. It processes locally — it works with no network. What is critical cannot depend on WiFi.
• Tracer — continuously maintains the carcass of origin → cut produced link. When a cut is recognized on the belt, it is associated with the carcass that was on the table at that moment (cross-referenced with the hook scan and the timing). If a food safety alert comes in for a slaughter batch, within minutes there is a complete list of affected cuts and their destinations.
• Labeling agent — cross-checks vision, scale weight, carcass of origin and commercial destination, and triggers the correct label on the weigh-labeler (Bizerba, Marel, Mettler Toledo). If something does not add up (the recognized cut is not the one selected manually, the weight is out of range for that category), the agent holds the label and the person validates. No label goes out until the system confirms that carcass, weight, category and destination all agree.

See the full IRIS architecture →

Before and after

Traceability by the operator's eye vs. traceability with iLEAN Vision

Aspect	Classic cutting + the operator's eye	With iLEAN Vision + Tracer + Agent
Recognizing the cut on the way out	Operator's eye + weigh-labeler	Continuous AI vision, without breaking the rhythm
Carcass → cut link	In the room supervisor's head	Recorded in Tracer at second zero
Mixed cuts on the same belt	Risk of mix-ups at peak hours	Cut-by-cut segmentation, whatever the order
The veteran's knowledge	Walks out the day they retire	Captured in the model, a permanent capability
Response to a batch food safety alert	Manual reconstruction, hours	List of affected cuts, in minutes
Mislabeling (carcass/category)	If it happens, the customer finds out	The agent holds the label before printing

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 cutting room. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic, reviewing your cutting manual, your line speed and your weigh-labeler.

• Mid-sized cutting room (pork 500-2,000 carcasses/day or beef 80-300 carcasses/day), a mainstream weigh-labeler (Bizerba, Marel, Mettler Toledo), a vertical meat-industry ERP, and growing retailer pressure for cut-by-cut traceability.
• Vision pilot on one outfeed line (camera + Edge + integration with the scale and the ERP). First value expected within a few weeks: recognition above 90% on the main cuts and the carcass → cut link maintained with no manual scanning.
• Indicative payback between 4 and 9 months, depending on cutting volume, average cut value, historical frequency of labeling incidents and how much operator time goes into manual scanning today.
• The hard lever: the day a food safety alert on a slaughter batch arrives, the difference between recalling exactly the affected cuts and recalling, out of an abundance of caution, the entire day's output — that alone pays for the pilot several times over.

And the plant manager's reasonable doubt

“What if vision mistakes a shoulder for a leg and it gets labeled wrong?” — hallucination is a problem of free generation, not of anchored tasks. Recognizing a specific meat cut from a closed catalog, trained on cuts from your room, is the anchored task par excellence — the best models brought error below 1.5% ^[1]. And even so: the agent never fires a label on its own if vision, weight and the manual selection do not agree — it holds, and the person validates. The three safety rings exist precisely for this.

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