Developed and delivered the Mooncake KV Cache Transfer Mechanism for the jd-opensource/xllm repository, enabling efficient inter-node data transfer and improved memory management in distributed systems with a focus on NPU integration. This C++ feature reduced transfer overhead and enhanced state sharing across processing units, laying the groundwork for higher throughput in distributed workloads. Additionally, contributed to the ping1jing2/sglang repository by improving the robustness of backend configuration loading using Python, implementing JSON decode error handling and explicit exception raising to prevent runtime failures. The work strengthened operational stability and maintainability for production backend deployments in both projects.