#cuda

An agent harness that compiles a model into one provably-correct, self-retargeting CUDA megakernel and self-tunes it past cuBLAS at batch-1 LLM decode, paper: https://arxiv.org/abs/2606.09682

Predictive GPU memory framework for AI inference workflows

Open hardware desktop AI node: 4× Tesla V100, 128GB HBM2, PCIe/NVLink topology and V-Core liquid/air cooling.

GPU profiling and optimization SKILLS with bundled MCP server

On/off switch for routing Claude Code through a local model instead of Anthropic Pro/Max subscription. No API keys, no cloud fallback—local mode either works or fails cleanly. (tldr; Route Claude Code through a local model via Distrobox, powered by llama.cpp and LiteLLM.)