Aseptic packaging with AI vision — a vial with a particle is not an incident, it is a recalled batch.

The problem

The inspector qualified under the pharmacopeia is essential — and is asking for help.

Visual inspection of injectable vials is one of the most closely watched points in the regulation: USP <790>, Ph. Eur. 2.9.20, the revised GMP Annex 1. And the reality for the pharma operations director is this:

1. The qualified visual inspector gets tired. Reviewing 60 vials a minute for a full shift is physically exhausting. Fatigue is a fact documented in the pharmacopeia, not an opinion. And where the eye tires, the defect slips through.
2. The SKU changeover is the critical moment. Amber vial, clear vial, viscous liquid, lyophilized product, a monoclonal antibody formulation — every SKU calls for its own optical criteria, and calibration is done against a bag of reference defects that ages.
3. Defect traceability is lost. When something is rejected, why was it rejected? At what point in the isolator did the particle get in? Was it glass, fiber, detergent residue? Without systematic capture, every incident is investigated from scratch.

The financial consequence is brutal: a single batch recalled for particulate contamination in an injectable product costs far more than the industrial cost — there is the brand cost, the report to the health authority, and the replacement batch the hospital is still waiting for. The classic system works 99.9% of the time — and that 0.1% is what reaches the pharmaceutical RASFF.

How it fits the IRIS system

iLEAN does not replace the inspector — it takes away the mechanical tedium and hands them the cases where their judgment counts.

The problem in aseptic packaging is not a lack of eyes: it is that the qualified human eye should be resolving borderline cases, not counting vials on a conveyor at 60 ppm until it fatigues. iLEAN acts as the putty that covers the capture window: it sees every vial leaving the isolator, does the mass screening and passes the qualified inspector the doubtful ones already classified, with the image and the batch history.

Vision sees every vial leaving the isolator. Edge runs inference locally without taking images out of the cleanroom. The agent cross-checks against the formulation, the isolator cleaning and the batch history. The person signs; the line does not restart on its own.

The three iLEAN pieces applied to aseptic packaging:

• iLEAN Vision + Edge — a machine vision terminal (CNN) mechanically integrated over the isolator outfeed. It reads level, particles (including glass, fibers and residues), seal integrity (stopper, aluminum cap, crimp) and the appearance of the lyophilized/liquid product. Inference runs locally: the image does not leave the cleanroom except for the dossier. It works without a network: if the plant loses WiFi, Edge keeps rejecting defective vials.
• iLEAN Connect — captures the batch formulation (prior CIP/SIP, autoclave parameters, stopper type, viscosity) and makes it available to the model. If a SKU changeover means retraining parameters, it happens when the order is loaded, not halfway through the shift.
• iLEAN Agent — cross-checks Vision's rejects against the formulation, the isolator cleaning, the primary vial batch and the sterilization autoclave curve. If it detects a pattern (a cluster of glass particles at a specific position on the conveyor), it holds the batch and alerts the Qualified Person. It composes the per-batch GMP dossier automatically: list of rejects, image of each one, probable cause, required signatures.

See the full IRIS architecture →

Before and after

Classic qualified inspection vs. inspection with iLEAN Vision + Edge + Agent

Aspect	Classic human visual inspection	With iLEAN Vision + Edge + Agent
Coverage	100% by inspector, with fatigue accumulating through the shift	100% by machine, doubtful ones to the qualified inspector
Glass particle detection	Depends on the inspector's condition and the lighting	Constant per SKU, recorded with an image
SKU changeover	Manual recalibration, risk during the transition	Specific model loaded when the order is launched
Defect traceability	Signed sheet with a manual tally	Image + probable cause + batch history per reject
Operation without a network	n/a	Edge keeps running inference on the panel's own power
GMP / Annex 1 dossier	Rebuild by hand, days	Generated per batch, signed, ready for the inspector

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.

• Aseptic line of 2-10 ml injectable vials, integrated isolator (Bosch/Optima/IMA), typical speed 50-120 ppm, multi-SKU with clear and amber vials.
• Deployment of iLEAN Vision over the isolator outfeed + Agent for the GMP dossier. First value expected within a few weeks: detection for the pilot SKU.
• Indicative payback between 4 and 9 months, depending on the average cost of your current rejects from manual inspection, the frequency of past recalls and the qualified inspector time recovered for higher-value tasks.
• Expected reduction of the inspector's visual fatigue and of inter-inspector variability ≥ 30%, based on the pattern seen in other sectors with Edge inspection on the line.
• The hard lever is a single recalled batch avoided. One recall of a sterile injectable pays for the entire pilot and leaves change to scale up.

And the technical director's / Qualified Person's reasonable doubt

“What if the AI misses a critical particle?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely classifies what it sees against a validated pattern (which is what Vision does with every vial), the best models brought error below 1.5% ^[1]. And even so, the iLEAN architecture does not decide alone: it holds the doubtful unit and the qualified inspector signs. The pharma quality standard is in the same order of magnitude as the automotive one — 25 PPM ^[2] — and it is reached with human-machine jidoka, not by eliminating the inspector.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks. ^[2] Symestic, automotive quality standard (25 PPM).