Client engagement · anonymised
Clinical Document AI for an Australian Healthcare Practice
An LLMOps engagement for an Australian healthcare practice: governed RAG over clinical documents cut review time by about 70% in eight weeks.
- Industry
- Healthcare
- Region
- Australia
- Focus
- Generative AI & LLMOps
Client engagement — the client is anonymised at their request; figures are as measured on the engagement.
Challenge
A specialist healthcare practice of 5–20 staff was spending a large share of clinical time reading specialist reports, referral letters and patient records to find the few facts each consultation needed. Review was entirely manual, there was no structured AI workflow, and early experiments with AI tools had no governance around prompts or outputs — a non-starter for clinical information.
Solution
The approach: a governed generative AI workflow that reads clinical documents, retrieves the passages that matter and presents them as structured summaries, with every prompt versioned and every release evaluated before clinicians see it. Retrieval-augmented generation (RAG) grounds each answer in the source documents, so clinicians can check where every statement came from.
Approach
- 1
We mapped the document workflow with the clinical team: which documents arrive, who reads them, and which facts each consultation actually needs.
- 2
We designed a RAG architecture using vector search, so summaries are grounded in the practice's own documents rather than the model's general knowledge.
- 3
We built a clinical document extraction pipeline in Python and FastAPI, orchestrated with LangGraph on Vertex AI, with results stored in PostgreSQL.
- 4
We put an evaluation and prompt-testing framework in place, so every prompt change is versioned and scored against real clinical documents before release.
- 5
We deployed the system securely in containers with Docker and added monitoring of answer quality and usage in production.
Frequently asked questions
How is patient information protected?
The system runs in a secured deployment with access controls, and summaries are grounded in the practice's own documents. Data handling was designed around the practice's privacy obligations from the first workshop.
Can clinicians check where a summary came from?
Yes. Retrieval-augmented generation links each statement to the passages it was drawn from, so a clinician can open the source in one step.
How do you stop a prompt change from making answers worse?
Every change is versioned and scored against an evaluation set of real clinical documents before release. A change that lowers quality does not ship.