Pharma blister control with AI — the half-empty cavity that comes back as a patient complaint.

The problem

The blister is right 99% of the time. The remaining 1% is a market recall.

The blister is where several sources of error converge, and they almost never show up in isolation:

1. Empty cavity or broken tablet — the operator cannot inspect every cavity at line speed; classic systems do it on samples.
2. The wrong tablet — a mix-up in the feed after an SKU changeover that was badly propagated, two active ingredients sharing a commercial color but differing in dose.
3. A pinhole or channel in the seal — invisible to the eye. It breaks the moisture barrier and the product degrades on the shelf.
4. Serialization / braille — the code is not linked to the right batch at changeover, the braille comes out of spec; the regulator and the customer both notice.

The quality manager knows it. So does the operator. And even so, on a running thermoformer line, the classic system works 99% of the time. That 1% is what reaches the EMA or the FDA — and in pharma, a market recall is work for an entire team for months.

How it fits the IRIS system

iLEAN does not replace QA — it hands QA data integrity by default.

The problem is not a lack of quality: it is that quality is measured on samples and that process information (the MES recipe, the granulate batch, the serialization, the SKU changeover) lives on islands and never reaches the camera or the batch record in time. iLEAN acts as the putty between the thermoformer, the MES, the serialization system and the QA room.

Edge sees every cavity before die-cutting. Connect reads the SKU, the batch and the serialization code wherever they live. The agent cross-references them with the approved blister layout, holds the batch if something does not add up, and builds the dossier with a timestamp. The person signs — never the other way round.

The three iLEAN pieces applied to blister control:

• Edge — a terminal with machine vision (CNN) over the thermoformer. It inspects every cavity: presence, color, geometry, tablet integrity, seal (with dedicated lighting where applicable), braille position and serialization code. It fires the actuator before die-cutting. It works with no network.
• Connect — captures the SKU, the granulate batch and the serialization code whether they come from the MES, the ERP or the track & trace system. And it captures what arrives from outside (a regulator alert, a customer spec change, a note from the aluminum supplier) at second zero.
• Agent — cross-references recipe + batch + image + approved blister layout + serialization. If there is a deviation, it holds the batch, alerts QA and builds the dossier with a timestamp, a hash and the user's signature (ALCOA+ data integrity). The person releases the batch; the line does not restart on its own.

See the full IRIS architecture →

Before and after

Classic sample-based inspection vs. 100% control with iLEAN

Aspect	Sampling + visual check	With iLEAN Edge + Connect + Agent
Coverage	1 blister in every N	100%, cavity by cavity
Wrong tablet / dose	Caught in sampling or in a complaint	Caught inline by color and geometry
Pinhole in the seal	Leak test on samples	Dedicated lighting + CNN
Serialization / braille	Verified on samples	Verified and linked to the batch by the agent
Dossier for EMA/FDA	Rebuilt by hand, weeks	Automatic per batch, with ALCOA+
Response time to a deviation	Days	Seconds

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.

• Oral solids plant with 2-3 blister lines, a mix of PVC/aluminum and aluminum/aluminum, 6-12 active SKUs.
• Edge pilot on the priority thermoformer (camera + lighting + actuator + integration with MES and serialization). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, dominated by the avoided cost of a single market recall and by lower scrap at batch changeover.
• Expected reduction in deviations detected in QA of ≥30% in the first quarter and ≥70% by month six. The regulatory lever (an ALCOA+ dossier by default) is one of the strongest in pharma — it is worth a great deal ahead of your next audit.

And QA's reasonable doubt

“What if the AI hallucinates and releases a bad batch?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes a piece of data (reading the cavity and comparing it with the batch SKU), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds and the person releases. The three safety rings — a hallmark of IRIS architectures — exist precisely for this, and they fit the spirit of Annex 11 on human oversight of computerized systems.

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