# Overshoot

> Overshoot optimizes inference for AI teams that run their own models, from engine settings down to GPU kernels. If the deployment does not reach the agreed target, you do not pay.

## Pages

- [The Performance Engineering Blog](https://www.overshoot.ai/blog/): Measured results from the inference deployments Overshoot optimizes, with the posts behind them.
- [Method](https://www.overshoot.ai/method/): How Overshoot optimizes an inference deployment: nine levers, each change measured on your hardware against a fixed baseline and checked for output drift.
- [Engagement](https://www.overshoot.ai/engagement/): How an Overshoot engagement works: one metric, a measured baseline and an agreed target. If the deployment does not reach the target, you do not pay.
- [Tools](https://www.overshoot.ai/tools/): overshoot-roofline, coming soon: a profiler that compares each part of a vLLM or SGLang step with its hardware limit and shows how long the GPU waits on the CPU.
- [Play](https://www.overshoot.ai/play/): Four short games against a vision model: Quick Draw, Quick Math, Reaction Time and Aim Trainer.

## Posts

- [The case for inference optimization](https://www.overshoot.ai/blog/case-for-inference-optimization/): Which factors determine the best achievable performance of a deployment, and why reaching it takes more than a general platform.
- [A custom attention kernel for Gemma 4 on the RTX PRO 6000](https://www.overshoot.ai/blog/gemma4-attention-kernel/): No version of FlashAttention runs Gemma 4’s attention over the whole prompt on the RTX PRO 6000, so we wrote our own kernel.
- [Caching less to serve more requests](https://www.overshoot.ai/blog/long-drafts-memory/): For long, mostly accepted drafts, vLLM’s default way of undoing rejected tokens uses most of each request’s memory.
- [Compounding GPU and CPU work on a small model](https://www.overshoot.ai/blog/gpu-cpu-compounding/): Large speedups come from a stack of optimizations, each one exposing the next bottleneck. Here the bottleneck moved between the GPU and the CPU.
- [TCP for server-to-GPU communication](https://www.overshoot.ai/blog/tcp-latency-optimization/): A 5 MB upload over a 31 ms round trip took 325 ms on a cold connection. Connection reuse and a larger initial congestion window cut it to 72 ms. Interactive simulations show each step.
- [Multimodal processor rewrite: 5× lower latency](https://www.overshoot.ai/blog/qwen-multimodal-processor/): CPU preprocessing took 428 ms at p90 for a 15-frame Qwen video before the GPU saw it. Our bit-identical rewrite of the processor takes 81 ms, 5× less.
- [Gemma 4: 20× lower vision latency under load from one fix](https://www.overshoot.ai/blog/gemma-4-batched-encoder/): vLLM runs Gemma 4’s vision tower one frame at a time. Batching it cuts p95 time to first token on video requests about 20× under load, with bit-identical output.
- [Day-zero Gemma 4 support](https://www.overshoot.ai/blog/gemma-4-day-zero/): Google released Gemma 4, a family of four open multimodal models, on April 2, 2026. All four ran on Overshoot’s earlier hosted API the same day.

## Contact

- [Book a call](https://calendly.com/overshootai/meet-with-overshoot)
- Email: founders@overshoot.ai
- [X](https://x.com/overshootai)
- [LinkedIn](https://www.linkedin.com/company/overshootai/)
- [Y Combinator profile](https://www.ycombinator.com/companies/overshoot)