The industrial bakery line with AI — from mixing to packing, without paper as the glue.

The problem

A bakery line is a chain of processes that do not speak the same language — and the translator is a notebook.

An industrial bakery line is really three factories chained together with incompatible logics: the mixer works in batches (one dough, one recipe, one time), the tunnel oven is a continuous process that is very expensive to start and to stop, and packing is discrete and multi-SKU (formats, bags, best-before dates, retailers). Between one link and the next, today, the glue is paper and the shift lead's memory:

1. The work order comes down on paper — or in a spreadsheet somebody printed out — and every link interprets it its own way. The recipe lives in an R&D folder; the real mixer settings live in the head of the dough foreman.
2. Proofing is controlled by looking at the dough — the chamber's temperature and humidity readings exist on a panel, but they are not recorded anywhere. If the batch turns out badly, nobody can reconstruct why.
3. The tunnel oven is a continuous black box — its zone temperatures and belt speed sit on the mimic display of a panel from the nineties with no data port. Every start-up and every stop costs gas, waste and an hour of transient running.
4. Quality control is a sample — a few pieces an hour, checked by eye, written on a sheet. A bake color drift can run for half a shift before anyone names it.
5. Packing is where another factory begins — format changeovers, batches, best-before dates, and a traceability file that gets rebuilt by hand the day the IFS/BRC auditor asks for it.

The result: every link keeps its data on its own island — or nowhere at all — and the only whole-line view is the one the shift lead carries in their head. The day that shift lead retires, the line loses its information system.

How it fits the IRIS system

iLEAN does not replace your machines or your ERP — it puts one piece at every link and joins them up.

The line's bottleneck is not any single machine: it is that mixing, the oven and packing do not share data, while the standards (IFS/BRC, batch traceability) demand that they be treated as one single process. iLEAN is the putty between links — without changing the mixer, without retrofitting the tunnel oven, without replacing the ERP or MES you already have. And it is modular: each piece is deployed separately, starting with whichever hurts most in your plant.

Connect captures the work order, the recipes and the machine panels where they live. Edge inspects 100% of the pieces at the oven exit. The Agents plan the oven as a continuous work center and assemble the batch file on their own. The person validates — the system releases nothing without a signature.

The iLEAN pieces applied to the complete bakery line:

• Connect — the data enters where it lives, not where it ought to live. It captures the work order and the recipe whether they come from the ERP, from a planning spreadsheet or from the paper that goes down to the floor. And it reads the panels of vintage machines with a camera — mixer, proofing chamber, tunnel oven — digitizing temperatures, times and alarms that today only exist on a display. No retrofit, no touching the PLC, with a timestamp and batch number from second zero.
• Edge — 100% piece inspection at the oven exit. Shape, color as an indicator of bake level and visible defects, piece by piece, at line speed. Where manual sampling checks a few pieces an hour, Edge checks all of them and turns a half-shift drift into an alert in minutes — with the result recorded by batch, ready for the file.
• Agents — the oven is planned as what it is: a continuous work center. The planning agent groups orders into long campaigns by temperature and format family, minimizing oven start-ups, stops and thermal ramps, and synchronizes the mixing and proofing batches backwards and the packing format changeovers forwards. In parallel, the traceability agent consolidates the batch file: raw materials, real process parameters, Edge's inspection and packing data. The human planner signs off the sequence; without a signature, nothing is released.

The Brain that orchestrates the agents lives in Central; critical data stays in the plant's ring 1, not in some generic cloud (see the IRIS architecture and the three safety rings).

See the full IRIS architecture →

Before and after

The line on paper and memory vs. the line with iLEAN

Aspect	Paper + spreadsheets + the shift lead's memory	With iLEAN Connect + Edge + Agents
Mixing and proofing data	On the machine panel, or nowhere	Captured by camera, with timestamp and batch
Tunnel oven start-ups and stops	Whatever comes out of fitting work orders by hand	Long campaigns by family — minimum start-ups and ramps
Quality control at the oven exit	A few pieces an hour, checked by eye	100% of pieces: shape, color, visible defects
Detecting a bake drift	Half a shift, when somebody names it	Minutes — automatic alert on color trend
Mixing ↔ oven ↔ packing synchronization	Carried in the shift lead's head	Sequence proposed by the agent, signed by the planner
Batch file for IFS/BRC	Rebuilt by hand when it is requested	Consolidated automatically, retrievable by batch in seconds

Impact estimate

Impact estimate for your line — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your line. It is there so the committee has an order of magnitude; we refine it during the diagnostic.

• Industrial bakery with one or several tunnel oven lines, 20-80 active references, a mixed machine base (some modern, some with no connectivity) and batch records fully or partly on paper.
• Deployment piece by piece: you start with the link that hurts most (typically inspection at the oven exit or panel capture) and extend to the next one once the first is already paying for itself. First value expected within a few weeks on one pilot line.
• Expected value levers (estimate to be validated): fewer oven start-ups and stops thanks to campaign-based planning — every transient avoided is gas and waste that is not burned —, 100% piece control at the oven exit instead of sampling, and an automatic batch file that turns the IFS/BRC traceability exercise from days into seconds.
• Indicative payback between 4 and 9 months per piece deployed (estimate to be validated), depending on the number of references, the oven's energy cost, your current waste rate and the administrative effort the batch file takes in your plant.

And production's reasonable doubt

“What if the AI gets a piece classification wrong, or proposes the wrong oven sequence?” — hallucination is a problem of free generation, not of anchored tasks. Classifying the shape and color of a piece against a reference pattern, or sequencing orders against known temperature and format constraints, are exactly anchored tasks: the AI recontextualizes a measured value using deterministic rules. In this kind of task, the best models brought error below 1.5% ^[1]. And even so, the person validates: the planner signs the sequence before it is released and Edge's rejection thresholds are set by quality, not by the model. The three safety rings are there for exactly this.

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