Show HN: Agentic CUDA Kernel Optimizer

An agentic CUDA kernel optimizer automatically generates and refines GPU kernel implementations by iteratively generating code, validating correctness against reference outputs, benchmarking performance, and using GPU profiling data to inform optimizations. The system, powered by LangGraph, can modify both kernel code and launch configurations while exploring different implementations, ultimately retaining the fastest validated version. It requires Python 3.12+, NVIDIA GPU with CUDA Toolkit, CMake 3.24+, C++17 compiler, and OpenAI API access.