Prefilled syringe autoinjector UDI-DI serialization — a corrupt code is not an error, it is a market recall.

The problem

The UDI does not break on paper — it breaks on the line, right before aggregation.

A prefilled syringe autoinjector that reaches the market clean is the cross-check of four data points living in four different systems:

1. The assigned UDI-DI — the unique device identifier that the corporate serialization system generated for this unit.
2. The printed UDI-PI — the batch, the expiry date and the serial number that the printer applied on the syringe body or on the autoinjector.
3. The readable DataMatrix — the actual image, with its contrast, its position and its content, exactly as it ends up after assembly and insertion into the autoinjector.
4. Aggregation to case and pallet — the SSCC and the reconciliation with the national medicines verification system.

The QA manager knows this, but cannot inspect every unit by eye. The line runs at 200-400 units per minute. The standard camera reads the code, but does not cross-check it against the serialization queue — it only says "I read it" or "I don't". The classic setup works 99.8% of the time. The remaining 0.2% is the one that triggers a recall when that aggregated case reaches the customer with two duplicate codes or a code that does not exist in the national system.

How it fits the IRIS system

iLEAN Tracer does not replace your serialization system — it seals the cracks between serialization, printing, vision and aggregation.

The corrupt UDI-DI problem is not a lack of information: it is information living on four islands that, at the critical moment (the unit passing under the camera), is not cross-checked in time. iLEAN Tracer acts as the filler that closes those gaps, without asking you to change SAP, the corporate serialization system, or the assembly line.

Tracer assigns UDI-DI per unit. Edge reads the DataMatrix before aggregation. Connect captures the serialization queue wherever it comes from. The agent cross-checks, holds if it does not add up, and prepares the dossier. The person signs — never the other way round.

The iLEAN pieces applied to UDI-DI serialization of prefilled syringe autoinjectors:

• Tracer (functional layer on top of Edge) — assigns the UDI-DI to each unit by consuming the queue from the corporate serialization system, and keeps the syringe-autoinjector-carton-case-pallet traceability alive without losing the chain of association. Every unit carries its number, and every number carries its evidence.
• Edge — terminal with machine vision (CNN) on the assembly line. It reads the printed DataMatrix in milliseconds, validates content, contrast and position, and fires the rejector if the unit does not match the next expected item in the queue. It works without a network. If the plant loses its connection to the corporate system, Edge keeps reading and rejecting with its local buffer — because what is critical cannot depend on connectivity.
• Connect — captures the serialization queue whether it comes from the modern corporate system, from a pharma-vertical MES, or from a CSV file the IT manager drops into a shared folder. And it also captures what arrives from outside (a notified body alert, a rule change in the national system) at second zero.
• Agent — cross-checks the assigned queue, the Edge visual reading, the shift history and the regulatory rules (FDA 21 CFR Part 11, EU MDR, UDI). If there is a deviation, it does not send an email at 10pm: it holds batch release and alerts the QA manager through whichever channel they use. The person validates and signs; release does not go out on its own.

See the full IRIS architecture →

Before and after

UDI verification with a standalone camera vs. cross-checked verification with iLEAN Tracer

Aspect	Standalone vision camera + checklist	With iLEAN Tracer + Edge + Agent
UDI-DI assignment	Corporate queue, not cross-checked on the line	Assigned per unit and cross-checked against the visual reading
DataMatrix verification	Read/no read — content not checked against the queue	Content, contrast and position cross-checked against the expected unit
Duplicate UDI detection	Shows up at aggregation, already in the case	Before aggregation, on the line, in milliseconds
Operation without a network	Camera keeps going, queue drifts out of sync	Edge keeps reading with a local buffer and reconciles on reconnection
Reconciliation with the national system	Rebuilt by hand when it fails	Automatic batch dossier, with a DataMatrix image per unit
Batch release	Signing without cross-checked evidence	Person signs on the complete dossier, AI proposes, holds if it does not add up

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.

• Combination device line at 200-400 units per minute, multi-SKU with several autoinjector presentations, aggregation to case and pallet with SSCC.
• Tracer pilot on one line (Edge camera over the DataMatrix + rejection actuator + integration with the corporate serialization system and the ERP). First value expected within a few weeks.
• Expected reduction in units mass-rejected due to desynchronization between queue and printing: ≥ 30% over the baseline incident rate of your line.
• Indicative payback between 4 and 9 months, depending on the frequency of documented incidents and the average cost of disaggregation and reprocessing at your customer.
• The hard lever is a single recall for a corrupt UDI avoided: product recovered from the hospital chain, reverse logistics, destruction, brand damage and a letter to the notified body. One recall pays for the pilot and the ones after it.

And the reasonable doubt of the pharma CAIO

"What if the AI gets it wrong and lets a badly serialized unit through?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI is limited to recontextualizing a data point from one system into another (reading the DataMatrix and comparing it against the expected queue), the best models brought the error below 1.5% ^[1]. And even then, what is critical is not decided alone: iLEAN Tracer holds the release and the person signs. 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.