Nojman is a consulting studio of one. We take on scoped engagements — an inference stack that needs to cost less, a training pipeline that needs to be reproducible, a cluster that needs an owner.
Start a conversationA backtest sweep that took too long on one machine, rebuilt as a Ray job on a Kubernetes cluster of consumer GPUs. What the setup looks like, what broke, and where the economics actually land versus renting datacenter cards.
A market maker is two systems with opposite needs — an execution layer that must never block and a signal layer that wants heavy compute. Why Elixir owns the order loop, why Ray owns the models, and how the two halves talk without poisoning each other.
KubeRay makes a Ray cluster on Kubernetes look easy in the demo. At real scale the head node, the object store, and GPU fragmentation push back. A field guide to the failure modes that show up between ten GPUs and a hundred, and where the money goes.