A medical assistant clinicians can check

An assistant clinicians can ask medical questions in plain language, and trust. Not a general chatbot with a medical skin, but a tool that answers from vetted sources and shows its work.

In medicine a fluent answer is a liability if it cannot be checked. Every claim needs a source, every source needs to be current, and the source has to actually say what the answer claims.

On top of that, health data needs stricter access than a typical app, and clinicians bring records in formats few tools read.

A calm chat with suggested questions to start from, pinned conversations, saved artifacts and a knowledge base each institution can curate. Patient cases keep context across conversations, and a patient timeline and case summary turn uploaded records into something readable.

Every citation is checked twice: once that the source exists and is active, and again that its excerpt supports the claim it is attached to. Bad citations are kept in the record but never shown. While the answer streams, a guard holds back a short tail of text so safety checks can catch problems before they reach the screen.

Uploaded documents are treated strictly as data, never as instructions to the model. The knowledge base ingests public references such as DailyMed and MedlinePlus, chunked and embedded for search, with a job to re-embed stale chunks when models change.

The project closed with a full handoff: a platform handbook, a route and code reference, the access rules, and the deployment and operations guide, so the client's team can run and extend it without us.