Clinical Evidence Navigator

Why retrieval, not recall

Asking a language model what the literature says is the single most dangerous thing you can do with one in a clinical context, because a fabricated citation is indistinguishable from a real one at a glance. This tool never asks the model to remember. It retrieves live from PubMed and ClinicalTrials.gov, then asks the model to synthesize only what was retrieved, and renders the answer with inline citations pointing back at the source.

What it is for

Clinical-grade answers to therapy questions in seconds instead of hours - the kind of question that comes up mid-review and currently costs a reviewer a detour into three databases. The retrieval layer and the synthesis layer are kept deliberately separate, so the sourcing can be audited independently of the wording.

Verification, not just instruction

Telling a model to cite only what it was given is a request, not a guarantee, so the answer is checked before it is served: every PMID and NCT number is matched against the records actually retrieved, and anything that does not match is rewritten as unverified and counted on screen. This matters more than it sounds. A fabricated identifier is correctly formatted and often resolves to a real paper - just the wrong one - which makes it the hardest class of error for a reader to catch. The same pattern applies anywhere a system has to be right about an identifier: order numbers, claim IDs, part numbers.

Where it stands

Built, working, and demonstrated to a senior director and upper management. Deployment was postponed - not for a technical reason - with potential future use in the appeals department, where cited literature retrieval is the daily job. It is the clearest demonstration of the approach I would bring to any clinical AI build: ground it, cite it, and let the reviewer check it.

Built with