Worked on optimizing the Triton backend mean operation within the FlagOpen/FlagGems repository, focusing on enhancing GPU performance and efficiency. The approach involved designing and implementing heuristic-based optimizations specifically for non-inner mean calculations, as well as introducing tile size heuristic functions to maximize GPU throughput. This work required a strong foundation in algorithm design, GPU programming, and performance optimization, all executed using Python. The feature was finalized by merging local changes into the remote repository, consolidating the improvements. The project demonstrated depth in GPU-focused performance tuning and careful application of heuristics to address computational bottlenecks in backend operations.