GroveAI
Case Study

Private AI Deployment for an NHS Foundation Trust

Westbridge NHS Foundation Trust needed AI-powered clinical note summarisation but could not send patient data to external servers. We deployed a fully on-premises solution that runs entirely within the Trust's own infrastructure.

Client: Westbridge NHS Foundation TrustIndustry: HealthcareDuration: 10 weeks

Results

Impact delivered

47 min

Time saved per clinician per shift

Reclaimed from note summarisation and search tasks

94%

Clinical accuracy

Validated by consultant-led blinded evaluation

Zero

Data leaving the network

Complete on-premises processing with no external calls

89%

Clinician adoption rate

Active daily usage within eight weeks of launch

The Challenge

What they faced

Westbridge NHS Foundation Trust serves a population of over 450,000 across three hospital sites. Clinicians were spending an average of 90 minutes per shift summarising and searching through patient notes — time that could be spent on direct patient care. The Trust had explored cloud AI tools but faced insurmountable barriers: strict NHS Data Security and Protection Toolkit requirements, Caldicott Guardian concerns about patient data leaving the network, and a board-level commitment that no patient information would ever be processed by third-party services. They needed the power of modern language models with the absolute guarantee of on-premises data sovereignty.

Our Solution

How we solved it

We deployed a locally hosted AI system running on the Trust's existing GPU-equipped servers. The solution uses a fine-tuned open-source language model (Llama 3.1 70B) optimised for clinical text comprehension. It integrates with the Trust's electronic patient record system to summarise clinical notes, highlight key findings, and generate discharge summary drafts. Every byte of data stays within the Trust's network — no external API calls, no cloud dependencies, no data ever leaves the building.

Approach

Step by step

01

Clinical workflow analysis

Shadowed clinicians across three wards to understand documentation workflows, pain points, and the specific types of summaries they needed.

02

Infrastructure assessment

Audited existing server capacity and identified two NVIDIA A100-equipped servers that could be repurposed for AI inference without additional hardware spend.

03

Model selection and fine-tuning

Evaluated seven open-source models against clinical text benchmarks. Selected Llama 3.1 70B and fine-tuned it using de-identified clinical notes provided by the Trust's research team.

04

EPR integration

Built secure integration with the Trust's SystmOne electronic patient record via HL7 FHIR, allowing clinicians to request summaries directly from the patient record interface.

05

Clinical validation and safety testing

Ran a blinded evaluation where consultants compared AI-generated summaries against manually written ones, achieving a 94% clinical accuracy rating.

06

Phased ward rollout

Deployed to the acute medical unit first, then expanded to surgical and paediatric wards over six weeks based on clinician feedback.

We were sceptical that AI could work within our constraints — no cloud, no external processing, full IG compliance. Grove proved it could. Our clinicians are getting time back with patients, and not a single byte of data has left our network. That matters enormously to us.

Dr Priya Mehta

Chief Clinical Information Officer, Westbridge NHS Foundation Trust

Technology

Stack used

Llama 3.1 70BvLLMNVIDIA A100HL7 FHIRPythonFastAPIDockerPostgreSQL

FAQ

Frequently asked questions

No. The entire system runs air-gapped within the Trust's network. Model weights, inference engine, and all application components are deployed on-premises. Updates are delivered via secure offline transfer.

Fine-tuning was performed on-site using de-identified clinical notes that had been through the Trust's formal anonymisation pipeline. The research team retained oversight throughout, and the process was approved by the Caldicott Guardian.

Every summary is clearly labelled as AI-generated and presented as a draft for clinical review. Clinicians can edit, approve, or discard summaries. The system also includes a feedback mechanism that flags recurring issues for model improvement.

Want similar results?

Book a free strategy call and we'll assess how AI can deliver impact for your business.