Worked on enhancing GPU-accelerated matrix multiplication in the facebookexperimental/triton and meta-pytorch/tritonbench repositories, focusing on Split-K GEMM autotuning and robust input handling. Leveraged Python, CUDA, and parallel computing to expand autotuning coverage, introduce deterministic reduction kernels, and optimize GPU utilization for undersaturated workloads. Addressed stability by implementing two-pass reduction strategies and input validation, preventing crashes from invalid or non-contiguous tensors. Improved production-path reliability by filtering out problematic configurations and ensuring correct execution of reduction steps. These efforts resulted in more reliable, scalable, and performant GEMM operations, while also streamlining benchmarking workflows and reducing maintenance overhead for machine learning workloads.