nCompass Technologies

nCompass sells AI agents that act as a performance engineer for AI platform teams, finding and fixing CPU and GPU bottlenecks in production deployments. The focus is generative recommenders, ranking systems and LLM serving stacks, where the vendor argues that custom stacks stay under optimized because they lack the open source communities that tune mainstream serving engines. The agents work across three layers: the CPU pipeline, including data movement, feature processing and synchronization points that leave the GPU waiting; serving and scheduling, including batching overlapped with GPU work; and kernels written for a specific architecture. The vendor reports results against reference implementations, such as 2.8 times lower end to end latency on the meta recsys generative recommenders repository, a custom attention kernel for HSTU, 24 percent more output token throughput than stock vLLM on Kimi K2.6, and 2.4 times the requests per second of stock SGLang on DeepSeek V4 Flash. It is aimed at teams that run models at scale and want the gains without hiring specialist performance engineers. The agents can also be used from an existing coding agent, with integrations listed for Claude Code and Copilot CLI. The founders are described as PhDs formerly at Qualcomm and Arm. Pricing is not published. The site directs visitors to book a 30 minute call with an engineer.