Worked on the ROCm/onnxruntime repository to enhance build stability and maintain correctness in MLAS convolution kernels, focusing on deep learning and machine learning infrastructure. Addressed a compilation warning in the convolution padding path by marking the pad_bottom variable as maybe_unused, ensuring that padding semantics remained consistent across architectures such as RISCV64. This targeted C++ fix reduced build noise without altering convolution behavior, supporting ongoing kernel maintenance and robust performance releases. The work improved maintainability and streamlined validation and release cycles for MLAS-based convolutions, demonstrating attention to low-level details and a methodical approach to cross-platform C++ development in machine learning systems.