Worked on GPU computing libraries, focusing on CUDA.jl and AMDGPU.jl, to improve memory management and FFT plan generation. Addressed variable-sized memory pinning in CUDA.jl by introducing a PinnedObject struct and refactoring pinning logic, which reduced memory-copy errors and increased robustness. Enhanced FFT workflows by refactoring plan generation functions for better type stability and region handling, applying constant propagation techniques, and updating tests to ensure correctness. These contributions, implemented in Julia with expertise in CUDA programming and performance optimization, strengthened the reliability and efficiency of GPU-accelerated computations, supporting more predictable and faster execution in real-world scientific workloads.