Insights
Writing on GPU-kernel correctness — the failure modes, the cost, and the evidence.
What silent GPU-kernel bugs actually cost
A silently-wrong kernel doesn't crash — it runs at scale. Here's the real cost: GPU-hours billed for broken work, untraceable quality regressions, and months of green CI hiding the failure.
The correctness illusion in LLM-generated GPU kernels
Why a single torch.allclose check accepts silently-wrong CUDA/Triton kernels as correct — and what a measured 26-op corpus revealed when we replaced it with an operator-aware oracle.
Stop shipping silently-wrong kernels
Open source, dual-licensed MIT / Apache-2.0. Validate your first kernel in five minutes — or talk to us about an enterprise pilot.