Case study · New York, USA · AI governance · channel partner
Arbis×
zkAGIArbis: every AI decision checked, inside a sealed vault
Arbis checks what AI agents are about to do, before they do it. We run those checks privately, and fast enough to sit in the loop.
- Client
- Arbis.ai, New York
- Delivered
- Private, fast AI checks
- Privacy
- Sealed hardware
- Status
- Live · US partner
The story in 60 seconds. Captions are on the video.
The problem
- Every agent step needs a check, so every second of delay adds up.
- The checks read customers' private data.
- Spending was hard to see and easy to run out of.
What we did
- Run every check inside locked, verified hardware.
- Stream replies word by word, as they are written.
- Daily budgets and alerts, so the account never runs dry.
The result
~70 ms
until the reply starts streaming
Every call
runs in verified, sealed hardware
Capped
daily budget with alerts
US
Arbis now resells zkAGI
Only with zkAGI
- The model provider never sees customer data in the clear.
- A record of where every decision was made.
- Cost controls built in, not an afterthought.
See it live
Links
Under the hood, for technical readers
- Mixture-of-agents on the TEE lane: a confidential VM on Phala, attestation checked on every call.
- Scrubbing is off for this workspace by policy; the enclave is the privacy boundary.
- Attestation verification runs off the critical path; judge mode cuts rule checks to one strong call.
Handling sensitive data on a hot path?
We will replay your workload and show you the results.
Talk to an engineerMore case studies