Case studies
Delivered engagements, published with the client's permission.
Governed AI for inline nanoparticle manufacturing
How a clinical-stage pharmaceutical company brought design-of-experiments discipline to a live inline homogenization campaign, without exposing identifiable proprietary data to a public AI model.
The challenge
The client had a proven batch process for empty PLGA nanoparticles and was moving to inline homogenization to support scale-up. New, downsized inline equipment was arriving within days, and the team intended to begin runs the moment it was plugged in. That created pressure to design well and design fast.
The process was genuinely hard to tune. Five active parameters were in play, and several were coupled rather than independent. Four quality attributes had to be satisfied at once: particle size, PDI, zeta potential, and laser Doppler velocimetry. The size target was a narrow band of roughly 550 to 650 nm inside a wider 400 to 800 nm specification.
One constraint was firm. The client had been deliberately cautious about public AI tools, and proprietary process data could not be sent to them. The team did not want a black box. They wanted experimental discipline they could trust, trace, and defend to their own security, procurement, and regulatory reviewers.
The approach
We did not propose a predictive machine-learning model. With inline runs only just beginning, the honest and higher-value deliverable was a structured experimental architecture: a graded constraint map organized by evidence type, a sequential design of experiments that adapts as evidence accumulates, and a documented method for updating both as inline data arrives.
An evidence-graded constraint map
Built as a graded priority map rather than a binary diagram. Green regions carried the strongest prior support and were tested first; red regions were hold-or-avoid zones set by equipment limits or safety. Every element carried an evidence-source tag. Nothing inferred was ever dressed up as something measured.
An adaptive design of experiments
Classical sequential DOE with expert-encoded constraints: interpretable to the client's scientists and compatible with standard process-development practice. A multi-response scoring rule was locked in early, so trade-offs were resolved by a pre-agreed rule rather than post-hoc judgment.
A secure, governed workflow
Real client data lived only inside a per-client secure workspace, encrypted at rest and in transit, with access restricted to named reviewers. Any data used with an external model was abstracted first. Controls were documented against NIST frameworks.
A live weekly operating model
Each week we pulled the latest inline data, refreshed the constraint map and DOE, and reviewed updated run recommendations in a standing working session. Every week produced an explicit versioned output.
The outcome
In six weeks, the client moved from a proven batch process and brand-new inline equipment to a disciplined, documented campaign.
Maximum information per run
A structured architecture meant each inline run sharpened the constraint map instead of producing isolated data points.
Trade-offs resolved by rule
Competing demands across size, PDI, and charge were settled by a pre-agreed scoring method, not after-the-fact debate.
A reusable, inspectable pipeline
The client can regenerate the constraint map as new inline data arrives and operate the pipeline independently.
Clean inputs to regulatory work
Evidence traceability was recorded at the point of decision, so the artifacts feed directly into downstream QbD and regulatory filings.
Zero IP exposure
Abstraction kept the client's IP from ever reaching an external model.
If a process development problem in your pipeline looks like this one, we can scope it in a call.