An AI agent helped write a QR kernel 232 times faster than PyTorch

Sankalp entered a GPU kernel contest and used an AI agent to do most of the optimization work, ending with a batched QR decomposition kernel that runs 232 times faster than PyTorch's built-in torch.geqrf. In concrete terms, the geometric mean across a range of matrix shapes dropped from about 419,000 microseconds to 1,805. That was good for 12th place out of 183 entrants, reached over more than 1,500 submissions across 14 days.

The interesting part is how the work split. Sankalp ran GPT-5.5 as Codex for the main optimization agent, writing and revising Triton kernels, with Claude Pro as an advisor to clarify concepts and Modal handling GPU profiling. The human set direction and the algorithm: a blocked Householder method with WY representation, which keeps the slow column-by-column work inside narrow panels and turns the bulk of the math into large matrix multiplications that map well onto tensor cores. Most of the effort, Sankalp writes, went into reducing panel overhead and making the update more GEMM-friendly, not chasing memory or raw compute.

This is what agentic performance work actually looks like right now. Not one clever prompt, but 1,500 iterations of a human-directed loop where the model handles the tedious kernel rewrites and the human supplies the numerical insight.

Why it matters

If you do performance engineering, this is a realistic template: pick the algorithm yourself, then let an agent grind through the implementation and profiling loop. The 232x number is real, but so are the 14 days and 1,500 tries it took to get there.