Contributed to the vllm-project/tpu-inference repository by developing the TPU Offloading Host Memory Kind Override feature, which allows users to customize the type of host memory used during TPU offloads. This enhancement introduced greater flexibility in memory management for tensor operations, supporting both pinned and unpinned host memory scenarios. The implementation included comprehensive automated tests to ensure correct behavior across different configurations, emphasizing reliability and maintainability. Leveraging skills in TPU programming, back end development, and testing, the work established a foundation for future performance optimizations and safer memory usage, all delivered using Python as the primary development language.