What it is
A retrieval-augmented generation engine for regulatory affairs teams. Structure-aware ingestion, hybrid dense and keyword retrieval with reranking, citation-grounded generation. 10,000+ indexed documents, two pilot pharmaceutical clients.
The part that was hard
Not the retrieval. The evaluation. A curated question-answer test set, retrieval precision and recall, and grounding checks, used to decide whether a change to chunking or embedding or reranking had actually improved anything. Most of what I know about verifying language model outputs I learned by discovering how easy it is to convince yourself you have improved a system.
What I would do differently
To writeOne honest paragraph. This is the paragraph that separates a research page from a founder page.