How We Would Get Diagnostic AI Out of Jupyter Notebooks and Into Hospitals

02 — Our approach

Our approach would start with data archaeology, not engineering. We would trace every version of the training dataset and expect to find preprocessing steps that changed in a notebook months earlier without version control. The model in production may have been trained on different image normalisation than the one in the lab. We would freeze the preprocessing, design a pipeline with versioned datasets and model artifacts, and add a gated release process where every model must pass A/B testing on anonymised real hospital data before deployment. For compliance, we would configure immutable audit logs for every training run, every inference, and every human override. The explainability requirement would be the hardest: regulators want to know why the model flagged a specific scan, so we would integrate saliency maps that highlight the regions influencing each prediction.