What is kernel-pilot?

KernelPilot is an automated system for optimizing GPU programs, specifically the low-level code routines called CUDA kernels that control how computations run on NVIDIA graphics cards. When engineers working on AI frameworks like PyTorch, vLLM, or SGLang want to make a particular math operation run faster on a GPU, they normally face a tedious cycle of writing code, profiling what is slow, researching prior solutions, editing, and repeating. KernelPilot automates much of that loop. It works through three cooperating components: an agent loop that takes a kernel definition and a performance target then plans, implements, tests, benchmarks, and profiles iterations autonomously, a knowledge base containing over 3,600 real merged pull requests from 14 upstream GPU computing repositories along with 52 wiki pages covering hardware techniques and kernel patterns, and a profiling skill that reads Nsight Compute reports (NVIDIA's GPU profiler) and translates raw metrics into a specific diagnosis and a single concrete next edit. The loop creates its optimization work in a clean standalone repository separate from the original framework, so experimentation does not pollute production codebases. It tracks provenance so engineers can trace which prior pull request or profiling finding influenced each change. A review gate using a system called Humanize RLCR prevents the loop from declaring success prematurely. This tool is for serious GPU kernel engineers who want to automate the research-profile-iterate cycle rather than do it manually. The primary language is Python.