KKH Chatbot

Nurses need clinical guidelines and dosage calculations at the bedside, under time pressure. Looking either one up manually is slow at exactly the moment speed matters most.

The system is a responsive web chatbot that answers guideline questions and runs medical calculators from plain language. It uses natural language processing and semantic search so a nurse can ask a question the way they would say it, rather than matching an exact document title.

I built the middleware that routes between the NLP engine and the clinical data, which is where correctness actually lives: the retrieval has to return the right guideline and the calculators have to execute exactly. The interface is mobile-first, because it is used standing up, and includes gamified quizzes and clinical scenarios for ongoing training.

Working against a clinical brief changed how I thought about failure. In most projects a wrong answer is a bug; here a plausible-but-wrong dosage is worse than no answer at all, so the calculators execute deterministically rather than being generated.

If I revisited it I would spend more time on evaluating retrieval quality systematically instead of testing by hand against questions I thought to ask.