Building Trust for Users and Regulators Alike: A Cost-Efficient PyTorch Pat... Raja Gopal Hari Vijay

This talk explores the concept of Compliance-as-Code, focusing on how it integrates regulatory controls into the training and inference pipelines using PyTorch. The speaker discusses the limitations of traditional compliance methods and introduces tools like Opacus, CrypTen, and Captum, while emphasizing the importance of end-to-end lifecycle traceability. Key topics include real-time auditing and compliance gates, showcasing features such as the Dispatcher, custom Autograd functions, and the hook system, which facilitate the embedding of logging and risk controls. The session highlights practical PyTorch patterns for creating automated documentation, immutable audit trails, and expediting certification processes in regulated AI environments.