Worked on optimizing local replica data transfers within the kvcache-ai/Mooncake repository, focusing on improving the efficiency and reliability of the transfer subsystem. Leveraged C++ to implement LOCAL_MEMCPY for metadata local replicas by passing the client’s local hostname into the transfer logic, enabling accurate detection of local memory copy scenarios. Adjusted the transfer engine to maintain correct endpoint matching for peer-to-peer paths while supporting metadata-local reads. Expanded test coverage using software development and testing skills, validating behavior across various transfer modes and hot-cache scenarios. These enhancements reduced local transfer latency and strengthened the robustness of data transfer operations in production.