Back to case studies Case Study

Major Teaching Hospital (ANZ)

Healthcare

40% faster policy answers

"Clinicians find protocols in seconds, not minutes."

Where we started

Clinicians toggled between twelve systems—EMR, pathology, radiology, SharePoint, and research registries—to answer seemingly simple questions. Discharge notes were recreated from scratch and the policy most recently emailed around was treated as the “source of truth”.

  • Pain

    4–5 hours/day on documentation, duplicate data entry, and no way to cite official protocols quickly.

  • Risk

    No audit trail proving which policy version informed a clinical decision.

What we built

Private knowledge assistant

Hybrid retrieval (BM25 + vector) with ACLs synced from Cerner roles and research groups. Answers include source paragraphs, timestamps, and a “show me the doc” link.

Note drafting copilots

Templates for discharge summaries, referral letters, and MDT notes auto-fill vitals, meds, labs, and imaging impressions. Clinicians edit and sign-off inside the EMR.

Evaluation + observability

130 SME-authored golden questions run nightly. We log groundedness, hallucinations, and citation accuracy; dashboards live in Power BI.

Governance controls

Azure AD SSO, PHI redaction at ingestion, and tamper-evident logs forwarded to Splunk. Every answer is tied to a clinician ID.

Six-week rollout

  1. Weeks 1–2

    Connected SharePoint, policy drives, and Cerner exports. Built the ingestion pipeline and defined SME golden questions.

  2. Weeks 3–4

    Shipped the knowledge assistant, evaluation dashboard, and the first discharge-summary template to a 12-doctor pilot.

  3. Weeks 5–6

    Scaled to 80 clinicians, added radiology + pathology connectors, and handed over runbooks plus on-call training.

Measured outcomes

2.5 hrs

Average time saved per clinician each day

95%

Responses rated “useful” or “very useful”

0

Security incidents or PHI breaches

Stack highlights

  • Azure OpenAI + open-weight fallback models hosted in the hospital’s subscription
  • Vector store + BM25 running on PostgreSQL with row-level security
  • Eval harness powered by LangSmith + nightly CI regressions
  • Teams and Cerner integration via Azure Functions + webhooks

Want to achieve similar results in your healthcare organisation?

Let's discuss your challenges