RiftStack

About RiftStack

RiftStack is an applied inference optimization lab. (This page was CloudRift's before August 2026, and the team, the research, and the benchmark history are unchanged.) We work on the software layer between models and hardware, so that the same model costs less to run and responds faster on the GPUs a team already has. Our main project is Emmy, an open-source ML compiler under Apache 2.0. Emmy lowers PyTorch through six inspectable intermediate representations down to CUDA, then searches for kernel rewrites tuned to the specific GPU in front of it. It ships as a drop-in vLLM plugin, and the code is at github.com/cloudrift-ai/emmy. On Gemma 4 12B in FP16 on an RTX 5090, Emmy's generated kernels reach up to 1.6x GEMM throughput over cuBLAS and a 1.30x geomean over PyTorch eager across 277 kernel shapes, with time to first token 16 percent below stock vLLM. The slowest shapes still trail eager. Our focus is consumer-class and commodity GPUs rather than top-end datacenter cards, because that is where most available supply sits and where compiler and serving choices make the biggest difference. The team builds low-level systems for performance-critical software, with backgrounds at Apple, Roblox, and Ubisoft. We are an NVIDIA Inception member and SOC 2 Type II certified. RiftStack is operated by CloudRift, Inc., which runs the sovereign AI platform at cloudrift.ai and production AI for a national telecom and bank. hello@riftstack.ai