This article details an auto-research contest where the author implemented batched square compact-Householder QR factorization using Codex in GPU Mode. Placing 12th out of 183 participants, they achieved a 232x speedup over the baseline solution within 14 days of over 1500 submissions. The problem required returning a compact Householder QR representation (H matrix with upper triangle R and tau vector) for batched square FP32 CUDA matrices.

Leveraging AgentGPT and the popcorn CLI, the author employed iterative loop engineering to push kernel optimizations through continuous testing and benchmarking. The competition centered on torch.geqrf reference implementation and utilized GPU Mode's popcorn CLI and modality credits system for agent-driven experimentation.