Case study · RAG · Web + mobile
A clinical AI assistant that answers like a consultant and cites like a librarian.
I spearheaded and built from scratch an all-in-one clinical assistant for a healthcare information publisher. Doctors ask in plain language and get one connected, evidence-backed answer in seconds, drawn only from the company's own content.
The problem
The company already had what doctors needed: drug monographs, disease information, patient cases, clinical news, special and conference reports, and education courses. But it was spread across separate sections, and keyword search failed whenever a clinician phrased a question loosely. A generic chatbot was out of the question. In clinical work, a confident hallucination is worse than no answer.
How an answer is built
Take a real-world style question: "67-year-old female, CrCl 10, on sitagliptin." The assistant reads the drug monograph and the disease information, connects them, and cites each point. Replay it below.
How this answer was found
- Parsed patient context: elderly · CrCl 10 mL/min · DPP-4 inhibitor
- Semantic search → Drug monograph: Sitagliptin
- Semantic search → Disease: Chronic kidney disease
- Connected renal-dosing guidance across both sources
A CrCl of 10 mL/min puts her in the severe renal impairment band, so sitagliptin needs a renal dose adjustment rather than the standard dose1. It's worth reviewing that alongside her overall CKD management and glycaemic targets2.
Illustrative recreation. Not clinical advice; brand and content removed.
- Understand the question: extract patient context (age, renal function, current medication) from free text.
- Retrieve with meaning: semantic search finds the right monograph and disease sections even when the wording doesn't match.
- Connect the sources into one answer in the tone of a consultant talking to a colleague.
- Cite every point with tappable cards that open the full monograph or article.
Two experiences, one engine
Mobile
It lives inside the company's flagship app, remembers recent conversations and preferences on the device, and adapts to 👍 / 👎 feedback.
Web
References expand inline next to a live view of how each answer was found, so clinicians can check the reasoning instead of taking it on faith.
Guardrails that make it trustworthy
- First-party only: every answer draws only on the publisher's own data.
- Visible provenance: each claim maps to a source the clinician can open.
- Transparent retrieval: the "how this answer was found" trace shows what was searched and why.
- Feedback loop: 👍 / 👎 signals flow back to improve relevance.
Outcome
Whether on a ward round or at a desk, clinicians get to a confident, evidence-backed answer in seconds, and the company's content became a conversational product instead of a library to dig through.