Worked on quantization and sampling enhancements for the ggml-org/llama.cpp and Mintplex-Labs/whisper.cpp repositories, focusing on GPU-accelerated machine learning optimization. Developed CUDA and Metal kernels to support efficient quantized-to-FP32/FP16 conversions, enabling broader hardware compatibility and improved inference performance for multiple quantization formats. Standardized quantization data paths across backends, simplifying maintenance and scalability. Additionally, implemented a top-nsigma sampling method in llama.cpp, providing refined global control over sampling parameters and more deterministic generation behavior. Leveraged C++, CUDA, and Metal Shading Language to deliver low-level performance improvements, advanced tensor operations, and flexible sampling techniques aligned with evolving project requirements.