Sankalp Shubham writes about evaluations, post-training and coding agents. His current owned biography describes work on frontier benchmarks and reinforcement-learning environments through Prime Intellect’s RL residency.[1]
Start with Auto-research with Codex →
GitHub 
X · dejavucoder Career
His current biography describes more than four years of industry experience, including backend engineering, applied-AI consulting and a founding AI-engineer role at an early-stage startup. Those prior employers are not named. His October 2026 update shifts the current description to benchmark experiments and RL environments at Prime Intellect.[1]
Contributions
Shubham’s QR-kernel account describes using Codex to propose implementations, run the contest’s correctness and timing checks and revise the code over many attempts. The target was a batched Householder QR decomposition, so a numerical test and benchmark supplied a concrete feedback signal. He emphasizes new ideas and diverse approaches when repeated local edits stop improving the result. The reported speedup concerns that contest’s baseline and workload, rather than general GPU acceleration.[2]
Tools, writing & talks
Writing, research & talks
Start here
Auto-research with Codexsankalp.bearblog.devA detailed account of kernel optimization through repeated correctness checks and performance measurement.
Read the problem and feedback-loop sections before the speedup result. GPU programming, linear algebra and benchmark knowledge help; reproducing the result requires the specified contest workload, hardware and correctness checks.
Author[2]A blog post detailing how the author used Codex as an autonomous research loop—iterating over 1500 submissions across 14 days—to optimize a batched Householder QR decomposition CUDA kernel, placing 12th of 183 in a GPU Mode contest with a 232x speedup over baseline. It covers the math, the feedback-loop setup that made the problem 'auto-research-able', and lessons on escaping local optima via idea diversity.
Catalog workOriginal source