Tablet coating with AI — the defect is not visible in the lab: it is visible in the pan.

The problem

The defect is born in the pan and discovered in the lab.

Tablet coating works 98% of the time. The remaining 2% is what wrecks the whole week's planning — an entire batch rejected, reworked, or worse: released to the market and returned.

The coating operator knows something is wrong before the system does. They can tell from how the spray falls, they feel it in how the pan turns, they see the tablets coming out "dull". But between the moment they notice it and the moment lab sampling confirms it — 120 minutes of process have gone by. And by then the pan has three batches behind it.

The three pieces of data you would need to cross-reference to anticipate the problem live in three different places:

1. The batch recipe — theoretical polymer viscosity, suspension %TS, nominal spray rate. In the MES, or in an R&D sheet the scale-up lead updated last week.
2. What is happening in the pan right now — product temperature, peristaltic pump speed, inlet air flow, atomization pressure. In the equipment's SCADA, with no cross-reference to the recipe.
3. What can be seen inside the pan — color, surface appearance, presence of agglomerates, segregation. The operator sees it through the sight glass every 20 minutes, when they can.

Three islands, each doing its own job well, none of them talking to the others. The defect lives in the joint between the three — and nobody was there listening in time.

How it fits the IRIS system

iLEAN does not replace your PAT — it puts a brain on top of what it already measures.

Coating control is not a problem of missing sensors. A modern plant already has NIR, Raman, thermocouples and SCADA. The problem is that nobody cross-references those signals with the batch recipe, with the R&D scale-up note and with the image inside the pan in the same second. iLEAN is the putty that seals that crack, without asking you to change your coating equipment or your PAT.

Edge sees inside the pan. Connect captures the recipe, the SCADA parameters and the scale-up note that was sitting in a spreadsheet. The agent cross-references them, anticipates the drift and holds the batch before discharge. The person signs — never the other way round.

The three iLEAN pieces applied to tablet coating:

• Edge — a terminal with machine vision (CNN) over the pan's sight glass or at the sampling point. It reads color uniformity, surface appearance, presence of agglomerates and picking in real time. If something drifts away from the learned pattern of a good batch, it triggers the hold before discharge. It works with no network.
• Connect — captures the recipe whether it comes from the vertical MES, from modern PAT or from the R&D scale-up spreadsheet. And it also captures what arrives from outside: the email from the polymer supplier warning of a viscosity change in the new batch, the internal note on the last deviation. All of it at second zero.
• Agent — cross-references the recipe, the SCADA parameters, the PAT signals and the pan image. If the real curve separates from the pattern of a correct batch, it does not write an email at 10 pm: it notifies the coating lead through their own channel, proposes the adjustment and waits for the signature. The person validates; the line does not restart on its own.

See the full IRIS architecture →

Before and after

Coating with after-the-fact sampling vs. cross-referenced coating with iLEAN

Aspect	Classic coating + isolated PAT	With iLEAN Edge + Connect + Agent
Defect detection	Lab sampling hours later	Inside the pan, in real time
Recipe change / scale-up	An R&D note that reaches the MES late	Connect captures the note; the agent applies it
Actual polymer viscosity	Assumed from the datasheet	Cross-referenced with the supplier's email for that batch
Picking / sticking	Seen in the lab after the batch	Edge detects it and holds before discharge
File for an FDA/EMA inspection	Rebuilt from 4 different systems	Automatic per-batch dossier, with images
Operation with no network	n/a	Edge keeps inspecting on the cabinet's own light

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 2-4 coating pans, a multi-SKU mix (enteric coating + cosmetic coating), a vertical MES living alongside scale-up notes in spreadsheets.
• Edge pilot on one pan (camera over the sight glass + integration with SCADA and MES). First value expected within a few weeks: the rate of batches rejected for film defects down by ≥ 30%.
• Indicative payback between 4 and 9 months, depending on the average cost of a rejected coating batch and the frequency of incidents documented over the last few years.
• The hard lever is a single batch avoided: product rescued, a pan cycle not repeated, a packaging slot freed. One batch pays for the pilot.

And the quality assurance manager's reasonable doubt

“What if the AI fails and releases a batch with picking?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the model merely recontextualizes (reading the pan image and comparing it against the learned pattern of a good batch), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds the batch and the person signs with their GMP credential. The three safety rings exist precisely for this. And the pharma quality standard that really matters — the order of 25 PPM demanded in automotive ^[2] — is only reached with complete capture of reality, not with an audit after the fact.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks. ^[2] Symestic — automotive quality standard on the order of 25 PPM, a cross-industry benchmark for industrial quality.