Autocomp is a portable, extensible framework for LLM-driven kernel optimization across tensor accelerators. Point it at a kernel, pick your hardware target, and Autocomp speeds it up, automatically.