During August 2025, Mengtian enhanced the pytorch/pytorch repository by developing internal observability tools for dynamic shape recompilation workflows. Focusing on backend development and debugging, Mengtian implemented a logging utility within MLHub to capture and surface insights related to dynamic shape issues, streamlining the debugging process for model optimization. The work included refining PGO insights content to improve clarity for developers, as well as creating utility functions to automate the insertion of debugging information. Using Python and leveraging machine learning infrastructure, Mengtian’s contributions improved maintainability and developer experience, emphasizing internal tooling and documentation rather than user-facing features or bug fixes.