Retrieval-grounded drafting of FDA submission documents

The problem

Regulatory submissions are assembled by hand from clinical study reports by expert medical writers. The work is slow, expensive and hard to scale — and every sentence has to be defensible against the underlying data, which rules out unconstrained generation.

What we built

1. Designed a multi-step LLM pipeline in LangChain that generates each section of the regulatory document from raw clinical trial data.
2. Used a Retrieval-Augmented Generation architecture to pull context from across multiple clinical study reports, keeping generated content factually anchored and internally coherent.
3. Wrote specialised prompts per section — study design, results, safety profile — so each one meets its own regulatory conventions and scientific register.
4. Orchestrated sequenced LLM calls for structured, long-form content generation that mirrors how an expert medical writer builds a document.
5. Ran the output through domain-expert validation loops to verify accuracy before it entered the submission workflow.