Generative AI

LLM-assisted intake summarisation for clinical teams

A clinician-in-the-loop assistant that drafts intake summaries from nurse-patient conversations, with strict scope boundaries.

Typical duration: 4-week design, 8-12 weeks to pilot, then a clinical pilot of 8-12 weeks
Team shape: 1 ML lead + 1 full-stack engineer + a clinical informatics liaison from your side

What good looks like

Documentation flow

Structured drafts in seconds, always edited by a clinician

Decision boundary

Triage and clinical decisions remain with the clinician, never the model

Evidence trail

Every draft logged with edits for the safety committee

The problem this addresses

Emergency departments and intake teams want to reduce documentation overhead without changing who makes clinical decisions. The honest constraint is that no model output can directly assign an Australasian Triage Scale category, and any system has to operate inside the existing clinical governance process from day one. The kind of engagement we take on is a transcription-and-summarisation tool that produces a structured draft for a clinician to edit, never a system that decides anything itself.

How we'd approach it

Audio is captured at the intake desk, transcribed on-prem, and passed to a hosted LLM with a tightly constrained prompt and a structured output schema (presenting complaint, history, observations, allergies). The model never sees patient identifiers, those are stripped before the LLM call and re-attached client-side. We typically prefer a hosted frontier model over a fine-tuned local one because instruction-following quality matters more than data-residency once de-identification is properly enforced, but that's a call to make with the safety committee, not against them. Every draft is logged with the clinician's edits, which gives the safety committee an evidence trail and creates a labelled corpus for future evaluation.

What we'd build

A pilot deployment running on hospital-managed devices at the intake desk, plus an evaluation harness the clinical informatics team can use to re-score historical drafts against new model versions before any change goes live. Triage category assignment, clinical decision support, and any automated write to the EMR are out of scope by design, the clinician copies the edited summary across.

Honest considerations

If your clinical governance process can't accommodate weekly prompt or model changes, the iteration loop is going to be slow and you should plan accordingly. If patient-facing consent and signage haven't been worked through, that workstream will take longer than the technical build, start it on day one, not week ten. And if the use case requires the model to make any clinical decision, this is the wrong engagement, we don't take that work on.

See all example engagements Talk to us