Worked on improving the maintainability of architecture-specific kernels in the uxlfoundation/oneDNN repository, focusing on the AArch64 JIT SVE convolution kernel. Refactored the kernel’s conditional logic by introducing a new ldr_imm_check helper function, which streamlined the code and enhanced readability without altering core functionality or performance. This change reduced cognitive complexity for future contributors and facilitated easier future optimizations in the JIT compilation path. The work was implemented in C++ with a strong emphasis on low-level programming and performance optimization, aligning with project standards while minimizing risk to existing features and ensuring no changes to the external API.