Client
Privacy-sensitive organization
Service
Implement
Status
Shipped
Stack
Local open-weight models · sensitivity router · eval harness
A privacy-sensitive org's on‑prem AI.
A private assistant and agent stack on the client's own hardware. A routing layer decides per request, by sensitivity, what runs locally and what may leave.
The brief
A privacy-sensitive organization wanted the leverage of modern AI, but some of its data could never go to an outside API. The usual answer is all-or-nothing: force everything local and give up the strongest models, or send everything out and give up the guarantee.
What we built
A routing layer that decides per request, by sensitivity, what runs where. Sensitive work stays on open-weight models served on their own hardware and never leaves the network. Non-sensitive tasks can opt into a stronger external model when it earns its keep. The client sets the policy, and every call is logged.
On-prem · your network
Your team
Routerby sensitivity
Local modelssensitive data stays here
non-sensitive · opt-in
External cloud APIs
What makes it hold up
An eval harness confirms the local models clear their bar, and re-confirms it as new models are released. We sized the hardware and quantized the models to run on it, so the guarantee holds on machines they own, not on a promise.
Where it landed
We handed it over documented, and their team runs and tunes the routing themselves. Sensitive data has a hard boundary, the audit trail shows every decision, and the strongest models stay in reach for the work that’s allowed to use them.
Sensitive dataLocalstays on your hardware
What leavesYour callpolicy-based routing
Every requestLoggedfull audit trail
Read another case
All work → 
