Project Lumen
Citation-grounded medical search for 1,200 clinicians.
The challenge
Helix Health's clinicians referenced 14 separate knowledge systems daily — drug interactions, clinical guidelines, internal protocols, peer-reviewed abstracts. Search returned garbage, and the legal team had vetoed the obvious LLM-only solution after a malpractice review.
Average time to answer was 11 seconds. 22% of queries were abandoned. Clinicians had started keeping printed cheat-sheets in their pockets.
What we built
A retrieval-augmented pipeline over 3.2 million clinical documents — drug compendia, guidelines, internal protocols, peer-reviewed abstracts. Embedded with a domain-tuned model and reranked with a custom learning-to-rank classifier.
Every answer is grounded in citations the user can click through to the source. We built a confidence indicator that quietly downgrades the response when the underlying retrieval is weak — and refuses to answer at all when no sufficient grounding exists. Deployed on HIPAA-compliant infrastructure with audit logging and per-user access controls.