Developed a high-performance NAX Split-K GEMM implementation for large-K matrix multiplications in the ml-explore/mlx repository, focusing on GPU programming and numerical computing. The work involved optimizing the Metal backend to maximize compute efficiency on Apple hardware, leveraging both C++ and Python to deliver robust benchmarking scripts for performance measurement and regression checks. By establishing clear benchmarking and backend pathways, the developer improved throughput for large matrix operations and provided better visibility into performance characteristics. This foundation supports future kernel optimizations and demonstrates a methodical approach to performance engineering, with collaborative contributions and a focus on sustained, measurable gains.