Worked on enhancing reliability and shape stability in PyTorch-based machine learning workflows, focusing on bug fixes within the pytorch/tensordict and pytorch/rl repositories. Addressed an inconsistency in the unravel_keys function to ensure compatibility with torch.compile, thereby reducing runtime errors and simplifying integration for downstream users. Improved the LineariseRewards transform to preserve the trailing dimension of reward tensors, preventing unintended shape changes during reward calculations. These contributions, implemented using Python and PyTorch, emphasized robust debugging and thorough testing practices. The work improved maintainability and compatibility in data processing pipelines, supporting more stable and predictable machine learning model development.