During May 2026, contributed to the kvcache-ai/Mooncake and ai-dynamo/dynamo repositories by enhancing Mooncake’s offload and integration capabilities within vLLM. Developed features enabling accurate disk-tier read support and improved classification of local disk replicas by exposing C++ predicates to Python, facilitating more reliable diagnostics and offload workflows. Implemented a new RPC server for disk-tier reads in the standalone mooncake_client and integrated MooncakeConnector into vLLM to support push-based parameter disaggregation for prefill requests. Leveraged expertise in C++, Python, and RPC to strengthen backend architecture, focusing on robust API development and seamless cross-language interoperability for large-scale workloads.