Solving Poker in Custom WebGPU Kernels Without a Tensor Library

I built an open-source poker solver running entirely in the browser using custom WebGPU kernels instead of a general tensor library like PyTorch. By leveraging advanced coding agents to generate and optimize these kernels against a trusted reference implementation, I achieved significant performance gains. This approach demonstrates that when code generation is cheap and verifiable, custom solutions can outperform general-purpose libraries, shifting how we think about software architecture in the AI era.