Hospital Services Assistant

The problem

In healthcare, the hard question is not what an assistant can answer but what it must refuse. A helpful-sounding sentence that hints at a diagnosis, or a slow answer to someone describing chest pain, is the failure that matters.

What I built

• A draft → check → regenerate once → withhold pipeline: every written answer is assembled in full and checked by code before a patient reads it; an answer that fails is replaced by a notice with patient services and the 112 emergency number.
• Emergency warning signs in ten categories get a fixed, reviewed answer before any model call; self-harm disclosures get fixed text written by people.
• Answers from eight bilingual hospital documents with citations linked by the server, plus six weeks of availability computed from a staff-maintained schedule.
• Appointment requests only: consent required, no clinical fields, no slot held, kept for at most seven days.
• A protected staff panel where AI summaries and reply drafts are labelled “draft — human review required”.

Key decisions

• Two directions, two calibrations

  The output check is narrow, because withholding a good answer is costly. The crisis check on the way in has no negation guard, because “I don’t want to hurt myself, but…” is still a disclosure.

• Measured, not assumed

  An adversarial set of 34 bilingual red-flag questions measures what the phrasing check catches and what it misses — and the misses are recorded with reasons instead of hidden.