UDI on medical devices with AI — a miscoded PI is a recall notice, not an incident.

The problem

UDI says every unit is unique — and the line works as if it were a batch.

The regulation is clear: every medical device carries its UDI (DI + PI) in the marking and in the national registry (GUDID for the FDA, EUDAMED under MDR). On paper, that is four GS1 Application Identifiers and one dossier per serial. On the shop floor, the three pain points we always find:

1. DPM laser marking on steel or titanium that a traditional reader gets right 85% of the time, so someone reworks the rest by hand at the end of the shift.
2. PI out of sync with the MES — the expiry date in the PI comes from the MES, the batch from the ERP, the serial from the laser itself. If one of them changes and nobody flags it, the UDI dossier breaks silently.
3. GUDID/EUDAMED updates handled manually by the regulatory affairs person, with spreadsheets and one AS2 upload a month. The day sales changes the packaging without telling anyone, the registered DI stops matching the engraving.

The result: the per-serial UDI dossier is after-the-fact reconstruction, not in-line capture. And that is exactly what an MDR/FDA audit asks about. The classic system (operator + printer + spreadsheet) works 99% of the time. The 1% is the recall notice.

How it fits the IRIS system

iLEAN does not add another PLM — it seals the gap between the engraving and the registry.

UDI lives in four places: the laser/printer, the production MES, the regulatory ERP and the repository (GUDID/EUDAMED). In practice the four are disconnected and traceability gets rebuilt by hand. iLEAN acts as the filler that stitches the four together without asking you to change PLM, MES or ERP.

Edge reads the engraving in line, Connect feeds MES/ERP/GUDID/EUDAMED, an agent assembles the per-serial dossier and holds the unit if something does not add up. The person signs off — always.

The three iLEAN pieces applied to UDI:

• Edge — a terminal with machine vision (CNN) and ISO/IEC 15415 + AIM DPM 29158 verification on laser marking, blister, label or RFID tag. It measures the quality grade, decodes GS1 Application Identifiers (01, 10, 17, 21…) and blocks the pack if the grade drops below threshold. Works with no network.
• Connect — capture from the laser, from the MES (batch, expiry date), from the ERP (DI references), from GUDID over AS2/SFTP and from EUDAMED through whichever channel is available. It also captures the regulatory manager's email (a packaging change from sales, a notice from the notified body) at second zero.
• Agent — it sustains the per-serial dossier. When the DI changes, it proposes the GUDID/EUDAMED update for a person to sign; when a PI does not fit, it holds and alerts. It also sustains the evidence pack for the auditor — the "boring, concrete agent" that ends the conversation.

See the full IRIS architecture →

Before and after

Manual UDI vs. UDI cross-checked with iLEAN

Aspect	Manual UDI + checklist	With iLEAN Edge + Connect + Agent
Reading the DPM marking	85-90% with a standard reader, manual rework	DPM-specific CNN, ISO grade measured in line
DI + PI composition	Left to the laser PLC	Validated in line against ERP and MES
GUDID/EUDAMED update	Manual monthly upload	Automatic proposal, human signature in one click
Per-serial dossier	Rebuilt for the auditor	Composed at second zero, always ready
Packaging change from sales	Reaches regulatory by luck	Connect captures it and the agent raises the alert
Operation with no network	n/a	Edge keeps verifying on cabinet power

Impact estimate

Impact estimate for your plant — to be validated with your own 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.

• Class II-III medical device plant (implants, instruments), one line with DPM laser marking and another with blister labelling.
• Edge pilot on the laser marking line, initial GUDID or EUDAMED integration from Connect, one agent for the per-serial dossier. First value expected in a few weeks.
• Indicative payback between 4 and 9 months, depending on the documented cost of rework from unreadable marking and of regulatory hours spent preparing an MDR/FDA audit.
• Hard lever: ≥ 30% reduction in marking rejects at QA and a noticeable drop in the time needed to prepare the per-serial dossier ahead of an audit.

And the reasonable doubt from the CAIO / regulatory affairs

"What if an agent publishes something to EUDAMED that was never approved?" — the three rings solve this by architecture. The agent lives in ring 3 with authority to propose; any external publication passes through ring 2 (validation) and only ring 1 launches the signed action. On reliability: hallucinations are a problem of free generation, not of anchored tasks such as cross-checking the PI read on the line against the GUDID record; the best models have brought the error rate below 1.5% ^[1]. It also complies with the human oversight logic the European AI Act requires of high-risk systems.

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

See also: all traceability solutions →