Developed a shared memory-aware filtering mechanism for matrix multiplication configurations in the FlagOpen/FlagGems repository, focusing on CUDA device constraints. The solution programmatically filtered out configurations that exceeded the device’s shared memory limits, directly addressing stability and performance issues during matmul planning on high-end GPUs. By implementing this targeted fix in Python and leveraging CUDA for performance optimization, the work prevented memory-overflow errors and improved runtime reliability. The approach demonstrated a strong understanding of CUDA memory management and matrix multiplication workflows, ensuring only feasible configurations were considered and supporting safer, faster production deployments without introducing new runtime failures or instability.