Regulatory document QC
Checking FDA filings in 30 minutes, not a working day
- Year
- 2024–25
- Client
- Johnson & Johnson
- Domain
- Pharma · FDA
- My role
- Lead AI/ML Engineer
- Stack
The problem
Investigator’s Brochures summarise hundreds of pages of nonclinical study data for the FDA. Medical writers were checking every number and claim by hand — 8 to 10 hours a document — and a single mismatch can trigger a clinical hold that costs millions a month.
What I built
- A retrieval pipeline that finds the exact study report, table and row behind every claim in the brochure.
- Study-ID synonym extraction (15+ domain patterns and compound aliases) that cut the search space by 70–90% and removed 40% of missed matches.
- A validation prompt that handles both numbers (with unit normalisation) and qualitative claims.
- Colour-coded Word annotations and a full trace from every verdict back to its source, so the output is audit-ready.
The hard part
Pharmaceutical tables break across pages — the header on one page, the data three pages later. I designed a parent-child embedding scheme that keeps rows linked to their headers, which lifted accuracy on complex PK, toxicology and genotoxicity tables by 20–30%.
Outcome
Review dropped from a working day to half an hour per document, errors fell to 1–2%, and the QA team went from 2 to 15 documents a week.
Client data and code are confidential; figures are from the delivered system.
Next caseAgentic credit analysis →