The same team that helped you decide takes it to a system that runs — the AI software, physical AI, or the infrastructure under it.
The work your people do by hand — answering the same questions, finding what’s buried in your systems, drafting, moving data between tools — done by a model instead.
Evaluated against your own baseline, and kept on infrastructure you control.
The compute your AI runs on when it can’t sit on someone else’s cloud — for sovereignty, data residency, latency, or because renting GPUs costs more than owning them.
Inference and training clusters, GPU, networking and storage — on-premise, colocated or air-gapped. Your data and models stay inside your walls.
AI that does physical work — a robot that sorts or handles parts, a line that inspects its own output, equipment that reads its own condition — running on the machine, not in a datacentre.
VLA and perception-plus-control policies on Jetson-class edge nodes, trained on your task until it holds in production. The process never pauses for a deployment window.
Three stages, and we own all three.
The same four stages across all three, but they mean different things — and you end up with different objects.
| Software & AI | Private AI | Physical AI | |
|---|---|---|---|
| Scope | Workflows, models, integrations and the acceptance criteria, fixed before code. | Cluster sized to your actual workload and budget — not to a datasheet. | The task, the environment, and the equipment the model has to run on. |
| Make | Built in increments on your real data, integrated as it goes. | Sourced under allocation and export rules, delivered, racked, configured. | Model trained, edge compute integrated with the machine and its controls. |
| Prove | Evaluated against the acceptance criteria on held-out real inputs. | Benchmarked on your workload before you sign it off. | Tested on the equipment, in the conditions it will actually work in. |
| Hand over | Deployed, documented, and in use by the people it was built for. | Commissioned and running, with the stack documented and your team trained. | Running on the machine, with the operations team trained on it. |
| You end up with | A system your people use by default. | Compute you own, with no hyperscaler dependency. | Equipment that sees, decides or acts without a person in the loop. |
Client names withheld under NDA; figures measured on the engagement.
Then don’t. The build is the expensive commitment — these are the small ones that make it safe. Each is a separate piece of work, each ends in something you keep, and none of them obliges you to build with us.
Earlier than that? The AI Readiness Score is free and takes three minutes. Weighing a cloud repatriation instead? That’s the Cloud Exit Assessment. · All decision products →
Bring the design if you have one — we’ll read it before quoting. If you’re not sure where it belongs, describe the problem and we’ll say.