API + excipient blending with AI — content uniformity is confirmed in the blender, not in the lab.

The problem

A hot spot in the bed is not seen by the probe. It is seen by the tablet.

Statistical sampling of the API + excipient blend works 99% of the time. It is the remaining 1% that wrecks the whole plan — content uniformity out of spec, batch rejected at compression, and if it travels further, a market recall for a potent API out of specification.

The classic problem is that the probe looks at some points of the bed — not all of them. If a hot spot forms exactly where no probe reaches (because the API agglomerated from static, because the density of the new batch changed, because the blender speed ran 30 seconds higher than usual), the blend passes the lab. And fails three stages later.

The three pieces of data you would need to cross-check to anticipate it live on three islands:

1. The batch recipe — nominal proportions of API and excipients, blender time and speed, lubrication instruction. In the MES or on a master manufacturing sheet.
2. What is happening in the blender — real speed, real time, real load, possibly in-line NIR if you have it. In the equipment's SCADA, with no cross-check against the recipe.
3. What today's raw material was like — real density of the API from the new batch, ambient humidity, excipient flowability. On the supplier's data sheet, in an email, in the warehouse lead's head.

Three islands, none of them talking to the others. The hot spot lives in the joint between the three — and nobody was there listening in time.

How it fits the IRIS system

iLEAN does not replace sampling — it adds vision over the whole bed.

Blend control is not a problem of missing protocols. The modern plant already has SOPs, probes and uniformity tests. The problem is that sampling is blind between points and nobody cross-checks the bed pattern with the recipe and with the real density of today's API. iLEAN is the putty that seals that crack.

Edge sees the whole bed in line. Connect captures the blender parameters and the raw material data sheet. The agent compares against the pattern of the good batch and, if the blend does not replicate it, holds. The person signs — never the other way round.

The three iLEAN pieces applied to API + excipient blending:

• Edge — a terminal with CNN vision over the blender sight glass (or an in-line NIR sensor if the customer already has the instrumentation). It reads the blending pattern, the presence of agglomerates and incipient segregation. It triggers a hold if the pattern does not match that of a correct batch. It works with no network.
• Connect — captures the recipe from the MES, the real parameters from the blender's SCADA (speed, time, load), the supplier's API batch data sheet (real density, particle size) and the emails that record a supplier change. All of it at second zero.
• Agent — cross-checks the blending pattern with the recipe, the CPPs and the raw material. If it detects that today's pattern diverges from the good batch (a hot spot that was not there before, over-blending caused by excessive speed), it does not write an email at 10 p.m.: it notifies the manufacturing lead and proposes stopping or redistributing before discharging to the next stage. The person signs; the line does not restart on its own.

See the full IRIS architecture →

Before and after

Blending with after-the-fact sampling vs. blending cross-checked with iLEAN

Aspect	Sampling + isolated SCADA	With iLEAN Edge + Connect + Agent
Hot spot detection	Blind between sampling points	Edge sees the whole bed in line
API density variation	Assumed from the data sheet	Connect captures the real batch data sheet
Over-blending / excess lubrication	Detected at compaction (batch already late)	Agent compares against the good-batch curve
Post-blend segregation	Seen in tablet uniformity	Edge detects it in transfer/hopper
File for inspection	Rebuilt from 4 systems	Per-batch dossier, automatic, with an image of the bed
Operation with no network	n/a	Edge keeps inspecting on the cabinet's own power

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 2-3 blenders (V-blender or bin blender), a multi-SKU mix with APIs of different densities, a vertical MES coexisting with paper batch notes.
• Edge pilot on one blender (sight glass + integration with SCADA and MES). First value expected within a few weeks: a reduction in batches rejected for content uniformity of ≥ 30%.
• Indicative payback between 4 and 9 months, depending on the average cost of a batch rejected for uniformity and the frequency of such incidents.
• The hard lever is one single batch avoided: product rescued, the compression stage not compromised, the packing slot freed up. One batch pays for the pilot.

And the quality assurance manager's reasonable doubt

"What if the AI releases a bad blend and it reaches compression?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the model merely compares the current blending pattern with the learned good-batch pattern, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds and the person signs with their GMP credential — the three safety rings exist precisely for this. And the cross-industry quality standard — on the order of the 25 PPM demanded in automotive ^[2] — is only reached with a complete capture of reality, not with after-the-fact sampling.

^[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 quality reference.