NorthwindRx

Client / Project

NorthwindRx is a global pharmaceutical company. The mandate came down from the Chief Scientific Officer: every regulatory affairs analyst should be able to ask plain-English questions of their internal corpus of FDA correspondence, clinical trial protocols, and historical submissions. The internal team had been trying for nine months.

Problem

The blocker was not technology. It was the gravitational pull of the procurement process — every vendor pitch turned into a 6-month POC that produced a slide deck, not software. The internal team had the talent but lacked the AI-native muscle memory to ship without a vendor. Meanwhile their analysts kept doing 4-hour manual searches through 47,000 documents.

Approach

A 4-engineer Project Pod embedded with two of their internal engineers for 12 weeks. We made one structural decision upfront: this would not be a vendor handoff — every line of code would be in their repo, every architecture decision would be reviewed with their team. They would own it on day one of production.

The system itself is conservative by design. RAG over Weaviate (their existing stack), Azure OpenAI for inference (their existing compliance posture), and the model never answers without citing the specific document and page number it drew from. Analysts can verify every answer against the source — non-negotiable for regulatory work.

Outcome

The assistant shipped to 200 regulatory affairs analysts in week 12 and reached 1,800 weekly active users by month four. Analyst median search time dropped from ~4 hours to under 15 minutes for the search patterns the system handles well. The internal team that paired with us now owns and extends the system; they shipped two follow-on features in the quarter after we left without external help.