Worked on quantization improvements for Hexagon NPU across the llama.cpp and ggml repositories, focusing on enhancing mixed-precision matrix multiplication accuracy and flexibility. Developed true Q8_0 quantization with configurable FP32 group sizes, integrating these options into the CMake build system to support production-ready tuning. Introduced inline optimizations and supporting utilities to the quantization path, improving both performance and maintainability. Ensured feature parity between llama.cpp and ggml for Hexagon NPU-based inference by aligning cross-repository enhancements. The work leveraged C programming, CMake, and embedded systems expertise, emphasizing performance optimization without addressing major bug fixes during the development period.