Granulation and compression in oral solids with AI — real RTRT, not a slide deck for the audit.

The problem

You have the PAT installed. You have ICH Q8/Q10 signed off. And you are still releasing batches on final testing.

The technical director of oral solids knows the script by heart: PAT, QbD, Design Space, RTRT, control strategy. The gap between the regulatory committee's slide deck and the reality of the tablet press is three to five years, and what is missing is not probes:

1. The PAT is installed, but the data lives in its own silo. NIR on one screen, Raman on another, press force sensors in the press SCADA, particle size somewhere else. Nobody cross-references any of it in real time with the recipe and the batch specifications.
2. The QbD model exists on paper but does not operate in line. The Design Space is characterized by the R&D team, but on the floor the decision still belongs to the operator watching three screens and calling the shift lead when something does not add up.
3. RTRT demands a signed release package, not “the press ran fine”. And rebuilding that package after the fact by cross-referencing SCADA, LIMS and PAT exports is exactly the manual work that keeps RTRT a goal rather than a reality.

The consequence is that the batch sits in quarantine for three to five days waiting for the QC lab's final test. Stock turnover suffers and the control model you paid for is not delivering. Not because the PAT is bad — because it is missing the brain that orchestrates it.

How it fits the IRIS system

iLEAN does not sell you a new QbD model — it operates the one you already have and assembles the signed release package.

The RTRT problem in oral solids is not theoretical: it is a problem of orchestrating data that already exists. NIR is there, Raman is there, the QbD model is there. What is missing is the putty that binds it all together in real time, cross-references it with recipe and specifications, and assembles the release package with a complete audit trail. iLEAN acts as that putty, without asking you to change probes, SCADA or the QbD model you validated with so much effort.

Edge captures the PAT in line. Vision reads the tablet press and the tablet itself. Brain cross-references it with the QbD model and assembles the RTRT package. The person signs; the press does not restart on its own.

The three iLEAN pieces applied to granulation and compression:

• iLEAN Edge — a terminal on the floor that ingests every PAT signal in streaming (NIR, Raman, FBRM, force sensors, moisture). It works with no network: if the WiFi drops, Edge keeps running the models locally and storing them with a signed audit trail.
• iLEAN Vision — machine vision over the tablet press (Korsch, Fette, Kikusui, IMA): tablet appearance control, sticking and picking, edge defects, embossing. And over the tableting belt: real-time classification of visual defects, before packaging.
• iLEAN Brain — the multi-agent brain that orchestrates everything. It cross-references the CQAs in real time against the Design Space you validated, decides whether the batch is in control, assembles the signed RTRT release package (audit trail, parameters, models, raw data, signatures) and hands it over to your quality system. If something leaves the Design Space, it holds the batch and alerts the responsible person.

See the full IRIS architecture →

Before and after

Classic compression with QC release vs. compression with iLEAN + RTRT

Aspect	Classic compression + release on final testing	With iLEAN Edge + Vision + Brain + RTRT
Time in quarantine	3-5 days waiting for analytical QC	Hours or same day, depending on the release plan
In-line PAT	Each probe on its own screen, cross-referenced afterwards	Unified streaming, cross-referenced with the recipe in real time
Tablet inspection	Manual sampling every N tablets	Vision over the belt, 100% of production
RTRT release package	Rebuilt by hand across SCADA + LIMS + PAT	Assembled automatically, signed, ready for the Qualified Person
Multi-API campaign changeover	Manual recalibration of screens and templates	Model reloaded per API, automatic audit trail
Final QC testing	Full panel per batch	Reduction ≥30% (estimate), full panel only for stability

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.

• Multi-API granulation/compression train (oral solids in blister), with PAT already installed (NIR + force sensors), targeting RTRT on 1-2 priority SKUs.
• Deployment of iLEAN Edge over the PAT + Vision over the press + Brain orchestrating the QbD model. First value expected within a few weeks: PAT-to-recipe cross-referencing in streaming on the pilot SKU.
• Indicative payback between 4 and 9 months, depending on your current analytical QC cost per batch, the average time in quarantine and the cost of stock-outs per SKU.
• Expected reduction of final tests per batch of ≥ 30% in the first year of RTRT — the Qualified Person signs every batch; what changes is what they have in front of them to sign.
• The hard lever is quarantine time recovered: faster stock turnover + less working capital tied up in quarantined product.

And the reasonable doubt of the QA director / Qualified Person

“What if the AI releases a batch that does not comply?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely verifies signals against a pre-validated Design Space (what Brain does with the QbD model your technical team approved), the best models brought error below 1.5% ^[1]. And even so, iLEAN does not release batches: it assembles the package and the Qualified Person 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 + well-orchestrated PAT.

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