These are drawn from my own work in regulated cell therapy development. Employers, products, and suppliers are not named, and details are generalized so nothing confidential is disclosed. What is described is what I did.
1. Right assay, wrong endpoint
The problem
A clinical assay run on samples shipped from multiple sites was producing too many results that could not be reported. The leading explanation, first surfaced by an AI tool and quickly adopted, was shipping temperature.
What I did
- Checked the hypothesis against data the organization already owned. The pre-transfer stability package had already tested temperatures beyond anything seen in transit, and performance held.
- Showed why the field data could not settle it: in passive shippers, long delays and rising temperature travel together, so the two were collinear. Partial correlation separated them; the temperature signal collapsed once hold time was accounted for.
- Designed a controlled experiment that varied temperature and hold time independently, so their effects could be separated by construction.
What it showed, and what changed
Hold time, not temperature, was the driver. The failure was evaluability: with delay, viability fell and debris rose until results could not be read. The transfer package had measured whether the analyte survived, not whether a result could still be reported. The assay was redesigned and revalidated, site handling was updated, and the failure rate went from chronic to a rare exception against a criterion set in advance. The investigation also led to a new SOP on how AI may be used in change control, as a supplemental tool with required human review.
2. Automated classification, taken to production
The problem
A new analytical capability needed automated classification of cell populations that expert analysts had always done by hand, without losing the rare populations that matter most.
What I did
- Took the capability from vendor selection through requirements, algorithm development, and validation into production.
- Designed the scoring the system optimized against: overlap with held-out expert ground truth, asymmetric penalties that protect rare critical classes over majority-class accuracy, and a separate throughput objective.
- Built the labeled ground truth, validated automated output against expert review, and fed misclassifications back into retraining as characterized, reproducible defects.
- Built the cloud data pipelines underneath: instrument data synced to structured storage, with downstream analysis triggered as each run landed.
3. A validated system, migrated
A validated document and quality system of record moved to cloud infrastructure. I revalidated it through the migration and preserved Part 11 controls, traceability, and audit trails throughout. Alongside that work I defined record attributes, required fields, and permitted values across the document and laboratory systems, so records stay classifiable and retrievable.
4. Standing up QC with no precedent
As a founding member of a Quality Control function for a novel iPSC-derived platform, there was no analytical precedent to build against. I developed and optimized flow cytometry release methods for clinical trial use, confirmed genomic integration and ruled out unwanted insertions or deletions with ddPCR and sequencing, and transferred methods between sites, training receiving teams and setting the criteria that cleared each site to operate.