Participated in public demo 2601 with GxRAG

The problem [00:23] In the EU, every pharmaceutical product requires an XEVMPD submission whenever something changes, a new product, a variation, a change in product insert, and teams have only 15 to 30 days to submit the update. Mo explains that today this process is largely manual: reading through 200+ pages of EMA guidance, cross-referencing hundreds of pages of regulatory dossiers, and populating 100+ structured data fields by hand, all against a strict compliance clock. What they built [01:11] GxRAG is an agentic RAG system built on three layers: a human-in-the-loop layer, an agentic workflow layer, and an audit/compliance layer that makes every AI decision traceable back through the chain from suggestion to human approval. In the live demo, Mo shows a user asking a regulatory question, how to transfer a marketing authorisation between organisations under the same headquarters, and GxRAG returns a cited, step-by-step answer, verified by an LLM judge across accuracy, completeness, and format, then lets the user extract the affected structured fields, each with a confidence score, source citation, and extraction notes. Fields move through a review-and-approval workflow, author submits, reviewer approves, approver signs off with an electronic signature (21 CFR Part 11-compliant), before export as JSON or XML for the next step in the pipeline. What happened [06:15] By the end of the sprint, Mo had solved three of his hardest technical problems: LLM-judge reliability (via YES/NO verdicts with HIGH/LOW confidence and rubric-based scoring), RAG chunk quality across 11,000+ chunks with complex tables and process maps (via vision-LLM extraction plus hybrid vector/BM25 retrieval with RRF/MMR reranking), and meeting GxP audit-trail requirements without sacrificing performance (via asynchronous, immutable audit capture). The ask [06:15] Mo is looking for architecture guidance: how to make the compliance/audit layer reusable across different RAG systems, rather than rebuilt bespoke for each one.