Immutable Infrastructure for Clinical AI Pipelines

This technical report discusses the importance of reproducibility in clinical AI systems, emphasizing the need for auditable, explainable, and secure AI models to comply with healthcare regulations such as FDA, MDR, and HIPAA. It outlines the challenges faced by traditional machine learning pipelines, including environment drift and version inconsistencies, which can lead to variable outcomes despite identical code. The report details how Bazel can be utilized to create deterministic and hermetic builds, ensuring consistent outputs across different environments. It describes the processes involved in using Bazel for preprocessing, model training, and deployment, highlighting the significance of fixed random seeds and sandboxing in achieving reproducibility. Additionally, the report presents the role of Software Bills of Materials (SBOMs) in maintaining transparency and compliance during audits. Overall, it emphasizes that reproducibility is foundational for building trust and ensuring patient safety in healthcare AI solutions.