Pharmacy & Healthcare DSSClinical Decision Support, NHS-scale design
Clinical software is judged on what happens when it's wrong. This was an exercise in designing the accountability layer before the models.
The problem
Drug-interaction, demand and disease prediction are all tractable ML problems. Deploying them into a clinical setting is not — you need to know which model version produced a recommendation, whether its inputs have drifted since training, and to reconstruct any decision months later for review.
The approach
RandomForest models for drug-interaction, demand and disease prediction served through a React and FastAPI stack, sitting on top of a model registry with explicit versioning, drift monitoring on the input distributions, and immutable audit logging of every prediction with its model version and features.
The outcome
An end-to-end reference design scoped to NHS-scale requirements, where the answer to "why did it say that, in March?" is a query rather than a shrug.