Allergens in infant food with AI — zero tolerance is not a slogan, it is a system.

The problem

Four realities, four systems, one single label — and an infant at the end of it.

In an infant food plant, a correct label is the intersection of four pieces of data that are almost never in the same place:

1. The batch recipe — exact to the gram, with supplier identification for every raw material (a single flour batch contaminated with gluten ruins the day). In the ERP, in the MES, or in the R&D manager's head.
2. Production order and CIP cleaning — what ran on the line before, which CIP cycle was applied, whether it reached the validated temperature and time, and which lab result confirms it. Data scattered across the SCADA, the quality record and the operator's sheet.
3. The image of the container at sealing — the actual label that goes to the shelf, the one a parent will read at two in the morning. Produced by packing, and not always checked against the recipe in real time.
4. The destination country's rules — EU 1169/2011 + 609/2013, but also post-Brexit UK, FSMA if you export to the US, GB internal market. The rules change, and not always when the catalogs get printed.

The quality manager knows this. The managing director knows that a recall in this sub-sector is not an incident — it is a crisis with almost guaranteed press coverage. The classic system works 99.9% of the time. That 0.1% is the one that makes the front page. And in baby food, that is not something you manage: it is something you prevent.

How it fits the IRIS system

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

The problem with allergen traceability in infant food is not a lack of procedures: it is information living on islands that, at the critical moment (the recipe changeover, the CIP cycle that falls short, the SKU change at packing), does not reach the person who has to decide in time. iLEAN acts as the putty that fills those gaps, without asking you to change your ERP, MES, SCADA or packing line.

Tracer holds batch and recipe. Edge sees the label and measures the CIP. Connect captures what arrives from outside. The agent cross-references it with the destination country's rules and, if something does not add up, holds the batch. The person signs.

The three iLEAN pieces applied to allergen traceability in infant food:

• iLEAN Tracer — granular batch traceability from the raw material (with its supplier certificate of analysis) to the jar on the shelf. It is not an extra ERP module; it is the layer that joins the physical milestones and the documentary milestones into a single auditable chain.
• iLEAN Edge (Vision) — terminals with machine vision (CNN) on the packing line: they read the actual label, extract the list of declared allergens and check it against the batch recipe. And they read the CIP sensors to confirm that cleaning reached the validated temperature and time. They work with no network — what is critical does not depend on WiFi.
• iLEAN Agent — cross-references recipe + CIP + label + destination country rules. If there is a deviation, it holds the batch before sealing and prepares the file the quality manager will read in order to decide. If a recall becomes necessary for an external reason, the same agent bounds the exact scope in minutes, not days.

See the full IRIS architecture →

Before and after

Classic traceability vs. cross-referenced traceability with iLEAN

Aspect	Classic procedure + audit	With iLEAN Tracer + Edge + Agent
Source of the recipe	ERP, MES or spreadsheet — the operator checks one of them	All sources unified at second zero
CIP validation between batches	A sheet signed by the operator	SCADA sensors + photo of the record + an agent that validates temperature and time
Detecting a recipe-label mismatch	A later audit, or a complaint	On the line, before sealing, in milliseconds
Destination country rules	A catalog updated every few months	A live rule per market; a retailer change enters at second zero
File for auditor / authority	Rebuilt by hand, weeks of work	Per-batch dossier, automatic, with the full chain
Speed of bounding a recall	Days — broad scope out of caution	Minutes — scope pinned to the exact batch

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.

• Infant food plant with multiple formats (jar, pouch, powder) and multiple recipes (with/without gluten, with/without dairy, with/without nuts), exporting to the EU + UK + possibly the US.
• Edge pilot on one critical packing line + iLEAN Tracer over batches from raw material receiving onward. First value expected within a few weeks: on-line holding of batches with a recipe-label mismatch.
• Indicative payback between 4 and 9 months, depending on the frequency of documented incidents, the average cost of a rework triggered by suspicion and your exposure to recall risk in regulated markets.
• A ≥ 30% reduction in the time needed to assemble the batch file for an auditor (FSSC 22000, customer brand, health authority) from the very first quarter. But that is not the hard lever: the hard lever is one avoided recall.

And the quality director's reasonable doubt

“What if the AI gets it wrong and lets a mislabeled batch through? This is infant food.” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI recontextualizes a piece of data (reading the label and comparing it with the recipe, validating that the CIP reached temperature and time), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds the batch, prepares the file and the person signs. The three safety rings are designed for exactly this — in baby food, human validation at sealing is not up for debate, it is guaranteed.

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