Subvisible particles in parenterals with AI — by the time you read it, the batch is already rejected.

The problem

The final measurement tells you whether the batch passes — the trouble is you already made it.

USP 788 is one of the most expensive controls to fail in parenterals. The average count of particles ≥10 µm and ≥25 µm decides whether the batch is released or rejected, and by the time the HIAC counter result arrives, the whole batch is already packed. If the average falls outside the limit, you find out late. The reasons repeat:

1. Slow drift nobody sees — the filter has been running for weeks and its retention capacity drops with every batch; the operator has no trend in front of them, only the figure from the latest test.
2. Multi-product with different thresholds — USP 788 for injectables, USP 789 for ophthalmics: if the wrong rule was applied, you find out during the audit.
3. Instrument data on an island — the HIAC spits out a PDF, somebody files it in a shared folder, and the data is never cross-referenced with the MES batch or with the product history.

The instrumentation is not what fails — the counter works perfectly. What fails is the information loop: the data exists but arrives late, with no trend and with no cross-reference to the rest of the plant. And that is the exact crack iLEAN seals.

How it fits the IRIS system

iLEAN does not replace the HIAC — it connects it to a brain that sees the trend.

The USP 788 promise is not about counting more particles; it is about seeing the drift before the official average goes out of range. The instrument is already there. What was missing was the librarian: someone to cross-reference every reading with the product, the batch, the filter history and the applicable rule, and to raise the flag with time to spare. That is what iLEAN does, without asking you to change the HIAC or the MES.

Edge reads the counter in every subgroup. Brain keeps the golden reference per product and the trend per filter. The agent alerts the manager before the average crosses the USP limit. The person decides.

The two iLEAN pieces applied to subvisible particles:

• Edge (instrument capture + vision where applicable) — a terminal in the room that reads the HIAC counter directly (serial port, USB, IP) or its export file, and normalizes every subgroup at second zero. For older equipment with no interface, a photo of the panel is recognized by vision and the data integrated. It works with no network: if the room loses WiFi, Edge keeps capturing and holding the data until the connection is back.
• Brain (trend engine + USP rules) — keeps the golden reference for each product/SKU and the trend per filter, per line and per shift. It applies the correct set of limits for the format (USP 788 injectable, USP 789 ophthalmic). As soon as the trend approaches the threshold, the agent alerts the production manager through whichever channel they use — before the full batch, not after.
• Dossier agent — writes the batch record with the particle curve, the filter in use, the actions taken and the manager's signature. Ready for inspection, with no need to rebuild the trail after the fact.

See the full IRIS architecture →

Before and after

USP 788 at the end of the batch vs. a live trend with iLEAN

Aspect	HIAC on an island + PDF in a folder	With iLEAN Edge + Brain
When the problem shows up	At the end, with the batch already packed	As soon as the trend drifts, before the limit
Rule applied (788/789)	Whichever the operator remembers	The one matching the SKU, automatically
Filter history	In the shift lead's head	Curve per filter, predictive alert on drift
Cross-reference with MES	Manual, at batch close	Automatic at second zero
Documentary trail for inspection	Rebuild it from PDFs and spreadsheets	Live batch dossier, signed, exportable
Multi-product / SKU changeover	Manual reset of thresholds	Brain switches golden reference and rule with the SKU

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.

• Parenteral plant with 2-4 lines, HIAC counters in the room, multi-product under USP 788/789, documented history of subvisible-particle rejections.
• Edge + Brain pilot on one line: HIAC capture + golden reference per product + drift alert per filter. First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the fully loaded cost of the batch and the frequency of subvisible-particle rejections.
• A ≥30% reduction in batch rejections due to subvisible particles. The hard lever is that every rejection avoided is days of line capacity recovered, especially on low-volume, high-margin SKUs.

And the quality manager's reasonable doubt

“What if the AI decides to release a batch it shouldn't?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely reads a curve from the instrument and applies a USP rule, the best models brought error below 1.5% ^[1]. And even so, iLEAN's three safety rings guarantee that batch release remains a human signature. iLEAN proposes, the quality manager releases. See the architecture.

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