Your infrastructure. Your data. Your agents.
For organizations whose data is not allowed to reach an external API. Open-weight models running on hardware you own, doing the same work, under the same operating discipline — with nothing leaving your network.
Why on-premises is worth the trouble
Open-weight models are now good enough to do this work without a cloud API behind them, and they run on hardware you can put in a rack or under a desk. Which model, and how much machine, is something we size to the workload during the audit — that answer changes as the models change, and it is not a decision you should have to inherit from a web page.
Nothing leaves the network.
Every model call is served locally. No external endpoints, no third-party inference, no vendor holding your prompts. Once it is deployed it can run fully air-gapped.
Open weights, no licence surprises.
We deploy models whose weights you can hold and whose licence your legal team reads once. No usage caps that move, no acceptable-use policy rewritten under you, no metering.
The log stays with you.
Every agent action is recorded on your side. Your team inspects, exports and retains it on your schedule, with your own tooling, without asking us for a copy.
Where this is the only version that works
Some organizations cannot send the work to somebody else’s machine. That is a data-residency question, and it is the only reason this deployment exists.
Healthcare.
Patient scheduling, clinical note prep, referral coordination, insurance follow-up — all running on infrastructure you already control.
Legal.
Document review, case timeline assembly, client intake, billing coordination. Work that touches privileged information without it ever leaving your systems.
Financial services.
Trade reconciliation, reporting, client communications, KYC workflows. Deployed inside your existing security perimeter rather than alongside it.
Government & defense.
Briefing prep, correspondence management, procurement tracking, scheduling. Air-gapped by design, not by workaround.
What you get
The engagement, plus the part that only applies on your metal.
Hardware sized to the workload.
We specify the machine against the work the audit found — how much runs concurrently, how long the documents are, how fast the answer has to come back. You buy it, we configure it, and you own it.
Models chosen, not defaulted.
Open-weight models picked for your workload and your box, and swapped when a better one lands. Which one it is today is the wrong thing to put on a website; it is the right thing to put in the proposal.
Agents and pipelines, per the map.
Exactly as on our own infrastructure: an agent where the work needs judgement every pass, a deterministic pipeline where the answer has to be the same every time. The hardware does not change that decision.
A dashboard you host.
Monitoring, task assignment and agent management, running on your side of the boundary. Full visibility into what every agent is doing, without a console we control.
Training and handoff.
Your team learns to run, tune and extend it. Documentation, runbooks, and direct access to our engineers while it is being stood up.
Ongoing operation, if you want it.
The monitoring and correction we do everywhere else, done here too — including model upgrades as better open weights arrive. Or take it in-house and run it yourself. It is your infrastructure either way.
Same product. Different address.
On-premises is a deployment choice about where your data is allowed to go. It is not a different engagement, and it does not buy you a lesser version of the work.
The audit comes first here too, and it produces the same three answers: this should be an agent, this should be a pipeline, this should be left alone.
Monitored in production, corrected when it drifts, reviewed on a schedule. The half of the work most vendors hand back to you is the half we keep.
Where inference happens, who owns the machine, and who holds the logs. That is the list. Everything else is the engagement you would get either way.
If your data can go to an external API, the managed deployment is simpler and we will say so.
Questions about running it on your own metal
Start with the audit.
A working session on your operations and your constraints, and a written map of what should be an agent, what should be a pipeline, what should be left alone — and what any of it needs from your hardware. You keep it either way.