How Overshoot launched on X

Overshoot tunes the inference stack an AI team already runs. It profiles that deployment against the hardware limit and closes the gap, down to a new GPU kernel when vLLM or SGLang has no fast path for the model on that card. The customer pays only if Overshoot meets the performance target. After months of this work inside production stacks, the company says those teams have cumulatively saved more than $5 million. The launch post calls the offering software built from those fixes. On the site, the work is still taken one deployment at a time. A published case is Gemma 4 31B on one RTX PRO 6000. FlashAttention has no kernel for the model's 512-dimension global attention heads on that GPU, so vLLM falls back to a general Triton kernel. On a 20,000-token prefill, that global attention step took 1,770 ms against a measured hardware floor of 294 ms. Overshoot's kernel brought it to 464 ms. Time to first token on the same prompt fell from 4.5 seconds on the default path to 2.3 seconds in the configuration they ship. Against an fp64 reference, the new kernel matched the stock kernel's mean and maximum error. Zakaria El Hjouji and Younes El Hjouji, cousins, founded the San Francisco company, part of Y Combinator's Winter 2026 batch. Zakaria, the CEO, previously wrote GPU kernels at Meta AI and built low-latency, high-throughput pricing systems at Uber. Younes, the CTO, was a computer vision frameworks engineer at Intel. He also wrote the Gemma 4 kernel note.