Worked on enhancing observability for the PyTorch Dynamo compilation pipeline by implementing unified logging of functorch configuration settings in both the pytorch/pytorch and pytorch/benchmark repositories. Leveraged Python for backend development, focusing on data logging and performance optimization to enable visibility into AOT Autograd settings and activation memory budgets across compile stages. Introduced feature-flag gating for safe rollout and rollback, and developed unit tests to validate logging accuracy and serialization. These improvements allowed for efficient querying of compile metrics, streamlined debugging, and facilitated data-driven tuning of AutoAC budgets, supporting scalable instrumentation for machine learning workloads in production environments.