Aseptic eye drops control with AI — sterility is not tested at the end: it is built into every minute of the fill.

The problem

Eye drops cannot be sterilized at the end: whatever happens in the room reaches the patient's eye.

Most multidose eye drops are packaged in an LDPE bottle with a dropper tip — a container that cannot survive terminal heat sterilization. That is why the process is aseptic filling: the solution is sterilized through a 0.22 µm membrane, and everything that happens downstream of that filter has to preserve sterility until the container is closed. The final sterility test samples a handful of units; the real guarantee is the process. And that process is a crossroads of things that almost never sit in the same system:

1. The recipe and the filler — from the MES, with filling parameters, capping torque and line speed.
2. The environment — a continuous particle counter (≥0.5 µm and ≥5 µm), air and surface sampling, differential pressures. Each with its own system or its own file.
3. The interventions — every time an operator reaches into the Grade A unidirectional airflow, the intervention has to be recorded and assessed. Today: the room logbook and someone's memory.
4. The tests — sterilizing filter integrity before and after the batch, container closure integrity (CCIT), a current media fill. Results on paper or in the LIMS, reviewed days later.

The sterile production manager knows it. The release review reconstructs after the fact what happened over hours of filling: if a particle excursion coincided with an intervention, the investigation is long and the decision — the whole batch or just a stretch of it? — is made late and on incomplete evidence. Annex 1 (2022) asks for exactly the opposite: a living Contamination Control Strategy with correlated data, not data archived on islands.

How it fits the IRIS system

iLEAN does not change the filler or the EM system — it puts a safety net under every minute of the fill.

The aseptic filling problem is not one of instrumentation: the room already measures almost everything. It is information on islands spread across the filler, the environmental monitoring system, the intervention logbook and the LIMS. iLEAN acts as the putty that fills those gaps without asking you to change your filler, your EM system or your validated process.

Edge watches the critical zone and every intervention. Connect reads particles, pressures, filter integrity and the recipe. The agent correlates in real time: excursion + intervention + the affected stretch of units. Stack light red if something does not add up. The person signs.

The iLEAN pieces applied to aseptic eye drops control:

• Edge — a terminal with machine vision (CNN) over the critical zone. It records every Grade A intervention with a timestamp and correlates it with the units under the filling point at that moment. It supervises tip and cap presence before closing, and the in-line CCIT station. It works with no network.
• Connect — continuously captures the particle counter and the air sampler, the differential pressures of the room cascade, the filter integrity test result (pre- and post-batch), the validity of the media fill and of personnel qualifications, and the MES recipe with its filling parameters and torque.
• Agent — watches the batch timeline in real time: is there a particle excursion while filling? Does it coincide with a recorded intervention? Did the filter pass its integrity test before start-up? Is the media fill current? If everything checks out, it records and moves on. If something does not add up, it holds the affected stretch and raises the alert. The responsible person assesses and signs. And at the end of the batch, the agent generates the ALCOA+ dossier ready for the release review.

See the full IRIS architecture →

Before and after

After-the-fact review vs. an aseptic fill correlated with iLEAN

Aspect	Room logbook + systems on islands	With iLEAN Edge + Connect + Agent
Particle excursion	Discovered during batch review	Pinned down in seconds to specific units
Grade A interventions	Logbook + the operator's memory	Timestamped record, correlated with the fill
0.22 µm filter integrity	Paper result, reviewed at closure	No valid pre-test, the line does not start
In-line CCIT	Rejects counted at the end of the batch	Every reject cross-checked with torque and components
Media fill and qualifications	Due dates tracked by hand	Alert and lockout when validity expires
File for the auditor	Manual multi-system rebuild	ALCOA+ dossier per batch, one click away

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.

• Ophthalmic plant with 1-3 aseptic filling lines (conventional in RABS/isolator, or BFS), a continuous environmental monitoring system, an MES or electronic batch record in place or in the pipeline.
• Edge pilot over the critical zone of 1 line + Connect over EM/MES/LIMS + a real-time correlation agent. First value expected within a few weeks.
• Indicative payback between 6 and 12 months, depending on the documented frequency of excursions and investigations and the average cost of an eye-drop batch in your portfolio.
• The hard lever is moving from scrapping whole batches for lack of evidence to holding only the affected stretch with correlated evidence + cutting the investigation time per excursion. A single batch rescued on solid evidence pays for the pilot.

And the sterile production director's reasonable doubt

“What if the AI misreads an intervention in Grade A?” — recording that a person reached into the unidirectional airflow, at what timestamp, and which units were under the filling point is an anchored task: detecting an event on video and aligning it with a counter. In anchored tasks, the best models bring error below 1.5% ^[1]. And even so, in sterile manufacturing what is critical is never decided alone: the agent holds the stretch and the person assesses and signs. The intervention assessment Annex 1 demands does not disappear — it gains the evidence it lacks today: the video, the timestamp and the particle counts from the same minute, on the same screen.

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