Worked on the fzyzcjy/Mooncake repository to address a critical issue in RDMA device selection during operation retries. Focused on improving the reliability of RDMA paths by ensuring that prioritized NIC ID lists are correctly respected based on the retry count, which reduced unpredictable behavior and support incidents. Utilized C++ and network programming skills to refine the retry logic, correcting both code quality issues and functional bugs related to RDMA device handling. Maintained clear documentation and traceability throughout the process, supporting future maintenance and onboarding. The work enhanced the stability and predictability of RDMA operations without introducing new features.