Developed and integrated float16 arithmetic support across both CUDA and CPU backends for the ggml-org/llama.cpp and Mintplex-Labs/whisper.cpp repositories, enabling low-precision computation for machine learning workloads. Focused on implementing pure float16 addition, subtraction, multiplication, and division operations, the work included introducing vectorized float16 functions and optimizing compute kernels to improve throughput and memory efficiency. Leveraging C, C++, and CUDA programming, the developer ensured backend consistency for FP16 operations, laying a foundation for broader low-precision deployment and streamlined future maintenance. The contributions addressed performance optimization and numerical computing challenges in large-scale ML systems without introducing new bugs.