Imperial’s UK Dementia Research Institute runs complex, regulated AI pipelines on sensitive clinical and home‑monitoring data, making “just use a managed cloud ML platform” a non‑starter. This talk presents a reference architecture for running Kubeflow, MLflow and Argo Workflows on on‑prem HPC to support dementia research, from raw sensor and imaging data through to trained models and reproducible analyses. It walks through how researchers submit experiments that span Slurm‑backed GPU/CPU nodes and Kubernetes‑hosted services, while keeping data within institutional control and traceable for ethics and governance review. Attendees will see practical patterns for experiment tracking, workflow orchestration, and model promotion that fit the realities of public‑sector research IT: constrained budgets, strict information‑governance controls, and long‑lived cohorts of vulnerable participants.