Worked on performance and debugging enhancements for SYCL-backed tensor operations in the ggml-org/llama.cpp and Mintplex-Labs/whisper.cpp repositories. Focused on backend development using C++ and SYCL, the work involved removing unnecessary f16 to f32 conversions during DNNL matrix multiplication to reduce memory copy overhead and improve runtime efficiency. Debugging utilities were refactored to return strings instead of void, enabling more flexible and detailed logging for tensor operations and copy processes. These changes streamlined computation, enhanced observability, and facilitated faster diagnosis and maintenance, contributing to more efficient high-performance computing workflows without introducing new bugs during the development period.