Zhiyuan Li focused on backend and performance engineering across AI inference repositories, delivering operator optimizations and architectural enhancements for ggerganov/llama.cpp and Mintplex-Labs/whisper.cpp. He unified naming conventions, expanded multi-core CPU and SYCL acceleration, and improved tensor operation handling using C, C++, and CUDA, which increased portability and throughput across AVX2, AVX512, and ARM architectures. In intel/intel-xpu-backend-for-triton, he addressed autotuning reliability by ensuring correct benchmarking configuration and removing deprecated parameters, resulting in more stable performance tuning. His work demonstrated depth in low-level programming, cross-hardware optimization, and maintainable code integration, supporting scalable AI workloads and developer productivity.