Worked on the ROCm/composable_kernel repository to enhance the performance and maintainability of batched GEMM operations in multi-dimensional GPU computing contexts. Focused on technical debt reduction, the work involved consolidating device structures and logic into shared files, reducing code duplication and streamlining future enhancements. By introducing common structs and reusable components, the developer improved consistency across WMMA and multi-dimensional GEMM variants, enabling safer and more efficient performance optimizations. Applied include-what-you-use principles to improve compilation hygiene. The project leveraged C++, CUDA, and high-performance computing techniques to establish a unified architecture that supports ongoing kernel interface stability and maintainability.