Pet food extrusion with AI — the SKU changeover is the most expensive moment of the day. And it hinges on an Excel file the operator never read.

The problem

Recipe change, SKU change, and the fine-tuning lives in the head of the operator whose shift just ended.

A premium pet food plant can run 30, 40 or 50 different SKUs in a week: kibble for adult dogs, kittens, seniors, special diets, different shapes (bone, star, cylinder), different sizes. Every SKU changeover means changing the recipe, adjusting steam flow at the preconditioner, extruder screw speed, die diameter, drying profile. That coordination is where the day falls apart.

It happens because the information lives in different places and different timeframes:

1. The formulator's recipe — in an Excel file owned by the R&D manager, updated when needed and sent by email or WhatsApp to the shift lead.
2. The preconditioner and extruder parameters — on the console, a screen with a history only the initiated can pull up.
3. The kibble reading — the operator eyeballs it at the dryer exit, decides whether it looks right, and often scraps the first minutes until the line stabilizes.
4. The retailer's specification — in another sheet, another system, sometimes only in the sales rep's head.

The SKU changeover is the most vulnerable moment of the day: many parameters to touch at once, little margin to validate, and if something is left badly adjusted, the first batch of the new SKU comes out off-shape — and that is tons going to rework or to bulk sale at a lower margin. The system works when the veteran shift lead is in. When they are not, it does not.

How it fits the IRIS system

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

The extrusion control problem is not a lack of information: it is information living in islands that, at the critical moment (the SKU changeover, the new recipe that came in without going through the shift lead), does not reach the decision-maker in time. iLEAN acts as the putty that fills those gaps, without asking you to change the extruder, the dryer or the formulator's Excel.

Edge sees the kibble at the dryer exit. Connect reads the recipe, the extruder console and the retailer's specification 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 pet food extrusion:

• Edge — a machine-vision (CNN) terminal over the belt at the dryer exit or before the coater. It reads kibble shape, size, color and spread in real time, without touching the line. If the batch drifts from the shape declared for the SKU, it fires an alert and proposes an adjustment to screw speed or steam flow. It also detects the end of the transient (when the line has stabilized) so good batch counting can start. It works without a network.
• Connect — captures the formulator's recipe whether it comes from an Excel file in a shared folder, a vertical MES or the ERP. It reads the extruder console and the preconditioner parameters at second zero, without waiting for the end-of-shift close. And it also captures what arrives from outside (the retailer's specification changed by email, the packing operator's complaint over WhatsApp) without anyone forwarding anything.
• Agent — cross-checks the recipe of the running SKU, the extruder parameters, Edge's reading and the history of previous SKU changeovers. If it detects that the adjustment will not stabilize before a threshold of lost tons, it does not send an email: it alerts the operator through whatever channel they use, with the concrete proposal. The operator validates and signs; the line does not adjust itself.

See the full IRIS architecture →

Before and after

Classic control vs. cross-checked control with iLEAN

Aspect	Classic control (operator + shift lead)	With iLEAN Edge + Connect + Agent
Recipe for the new SKU	Formulator's Excel, arrives by email/WhatsApp	Captured at second zero + cross-checked with the console
Kibble shape and size	Operator's eyeball, judgment call	Continuous in-line reading, on every batch
SKU changeover transient	First minutes by eye, sometimes minutes too many	End of transient detected as soon as it stabilizes
Final moisture	Sample taken at packing, already late	Anticipated by preconditioner + Edge correlation
Off-shape batch	Confirmed at packing, goes to rework	Redirected to bulk or adjusted before packing
Batch file for the retailer	Rebuilt by hand, weeks	Automatic per-batch dossier, with Edge photo

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.

• Premium pet food plant with 2 extruders, 40-60 active SKUs, a single dryer and a fat/palatant coater.
• Edge pilot on the belt at the dryer exit (vision camera + integration with the extruder console and the formulator's Excel). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the number of SKU changeovers per week and the tons lost to badly adjusted transients.
• Hard lever: ≥ 30% reduction of the transient batch that goes to rework or to bulk. Every well-adjusted SKU changeover is tons that come in at full margin instead of reduced margin.

And the production manager's reasonable doubt

"What if the agent proposes the wrong screw adjustment and I jam the extruder?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI correlates what is already measured (recipe, extruder parameters, Edge reading, SKU history), the best models brought error below 1.5% ^[1]. And even so, the AI does not decide the adjustment: the agent proposes, the operator signs. The three safety rings are there precisely for this.

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

AI sunflower oil refining → · AI rice whitening control →