During March 2025, Wang focused on backend development and performance optimization for the kvcache-ai/ktransformers repository. He refactored the CPUInfer backend initialization process in Python, introducing a lazy initialization strategy that activates only when the requested thread count increases. This approach eliminated repeated backend instantiation, reducing startup latency and improving throughput for concurrent CPU inference workloads. By adding lifecycle guards, Wang ensured the backend’s state remained consistent with thread requirements, minimizing potential race conditions. His work addressed resource utilization and scalability challenges in multi-threaded environments, demonstrating a thoughtful application of backend engineering principles and performance-focused design within the project.