Worked on optimizing numeric conversion paths in the kvcache-ai/sglang repository, focusing on improving the performance of BF16 to FP8 conversions. Applied expertise in C++ and low-level programming to simplify bit manipulation logic, resulting in a more efficient CPU-side conversion process. Leveraged CPU optimization techniques to reduce both processing latency and CPU cycles, directly enhancing throughput for AI model caching workloads. Maintained code quality and traceability by referencing relevant issues in commit messages. The work demonstrated a strong understanding of bit manipulation and performance engineering, delivering a targeted feature that improved the efficiency of numeric data handling in AI infrastructure.