Developed performance benchmarking and cache policy improvements across the kvcache-ai/Mooncake and jeejeelee/vllm repositories. Built a dedicated benchmarking suite in C++ and CMake to evaluate the BatchEvict workflow in Mooncake, integrating it with MasterService for internal access and updating CI pipelines to refine coverage reporting. In vllm, enhanced the CPU offload cache policy by propagating request context through touch operations, ensuring correctness under concurrent workloads. Added regression tests in Python to validate context handling, improving reliability and consistency of cache behavior. The work emphasized backend development, performance benchmarking, and robust software architecture to support scalable, production-grade systems.