Worked on the ROCm/aiter repository to enhance the stability and reliability of Triton kernel operations on AMD GPUs, focusing on GEMM and attention workloads. Addressed a critical bug by implementing LLVM backend hardening for the gfx950 architecture, specifically forcing num_stages=1 in essential Triton kernel loops to prevent software-pipeliner crashes and numeric miscompiles. This targeted adjustment stabilized the release_tmp3 build and improved numerical correctness across various GEMM and MHA configurations. Leveraged expertise in compiler optimization, GPU programming, and Python to deliver production-ready improvements, ensuring consistent performance and mitigating hardware-specific issues in machine learning workloads on AMD platforms.