Good private AI depends on people who own the hardware, understand the network, and can run the models without sending their work to someone else’s cloud. Sightline Works exists to make that normal — for teams that need real AI and do not have an infrastructure department to stand it up. We build and operate on-premises AI on NVIDIA GPUs we own: serving open-weight models, evaluating them on real workloads, and turning what holds up into software we install on machines the customer owns. The question is simple: is it useful on infrastructure you control, or only in a demo.
Sightline Compute is our research and development lab, and this Hub account is its public home. On GPUs and an RDMA fabric we own, we work on long-context serving, quantization, multi-model routing, sandboxed agents, and the network around them — segmentation, egress control, and what inference traffic actually looks like on the wire. We also evaluate new open-weight releases against real workloads, then fine-tune and adapt them: the same training and data pipelines we run here are what we stand up on a customer’s hardware so their documents, terminology, and field never leave the building.
What we are building toward is a fleet that can test a new model against its own work and stage its own refresh, detection and egress content that ships with the product, and reference builds from a serious workstation up to datacenter-class NVIDIA systems. Public models, datasets, and write-ups from that work will land in this org first.