Selected Work

Case-study summaries focused on challenge, role, and measurable outcomes, without disclosing confidential implementation detail.

Roche / GOSH partnership: multi-workstream applied AI leadership

Problem: move several healthcare AI opportunities from concept into clinically useful delivery.

Challenge: regulated setting, interoperability constraints, and diverse stakeholders across clinical, product, and technical teams.

My role: led cross-functional workstreams across prioritisation, technical direction, and delivery execution.

Outcome: translated strategy into implementable roadmaps across multiple workstreams and maintained delivery momentum with clinical, product, and technical stakeholders.

Public reference: Roche and Great Ormond Street Hospital.

Genomics document intelligence: PDF to FHIR to clinician dashboard

Problem: genomics signal was locked inside unstructured PDF reports, slowing downstream use.

Challenge: heterogeneous clinical language, mapping ambiguity, and interoperability requirements for safe reuse.

My role: built the NLP pipeline from extraction and normalization through FHIR mapping and dashboard integration.

Outcome: 80% reduction in manual review time and a practical pathway from report ingestion to clinician-facing decision support.

Public reference: FHIR standard (HL7 FHIR).

CKD progression modelling and clinical deployment pathway

Problem: support earlier intervention planning for children at risk of CKD progression.

Challenge: longitudinal clinical data complexity, interpretability requirements, and adoption pathway constraints.

My role: co-led modelling and deployment planning with clinical collaborators.

Outcome: retrospective validation on 692 children; best model XGBoost with F1 0.72 and ROC AUC 0.80; SHAP used for interpretability.

Public reference: publication list at Google Scholar profile.

Enterprise compliance RAG / LLM assistant

Problem: internal teams needed faster and more reliable answers to compliance and policy questions.

Challenge: balancing LLM usefulness with source traceability, governance, and user trust.

My role: led architecture and delivery of a retrieval-augmented assistant and operating controls.

Outcome: deployed an enterprise capability aimed at approximately 1,500 users with source traceability, governance controls, and practical guardrails.

Public reference: practical RAG patterns (LangChain RAG overview).

Cancer Wait Times automation and MS prescribing analytics app

Problem: key operational and prescribing insights relied on manual reporting cycles.

Challenge: fragmented datasets, naming inconsistencies, and dependence on manual reconciliation.

My role: led analytics automation and delivery of a Streamlit app using joins, fuzzy matching, and business-rule reconciliation.

Outcome: improved repeatability and reporting turnaround for service and clinical stakeholders by replacing manual reconciliation with automated joins, fuzzy matching, and rule-based workflows.

Public reference: Streamlit and NHS waiting times context (NHS England RTT statistics).