Per-tenant policy
Each tenant names its models in order. A risk team can approve a model once and know no agent will call another.
Platform
Model Router
Your models, your region, your policy. The router picks the model for every agent step from a policy you set per tenant, fails over when a model has trouble and tracks the cost of each decision.
Policy per tenant
You decide which models may touch your customers.
Model choice is a governance decision, so it lives in policy with the rest of your controls.
Cost per decision
Model spend is metered for every step, so you can see what each approved outcome actually cost.
Open-weight models in your region
Run open-weight models you can host where your data must stay, instead of sending every call to one vendor.
Fallback chains
One model has a bad hour. Your agents do not.
Customers on a live call should never hear an outage. The router handles it in the order you chose.
Try the first choice
Each call goes to the first model in your tenant policy.
Retry the blips
Rate limits, timeouts and server errors are retried with a short backoff.
Fail over in order
If a model keeps failing, the router moves to the next one in your chain. A model with no key set is skipped.
Record the route
Which model answered and whether a fallback was used goes on the audit ledger.
Your region
Keep data where your regulator expects it.
Spine is built so the cloud is a deployment choice. Model placement follows your residency needs across the US, Canada and India.
United StatesCanadaIndia
What runs today
Voice AI runs on Google Cloud in Mumbai (asia-south1) with open-weight models, designed for sub-second turns. Other placements are set up per customer.
Questions
The model router, answered.
No. The router speaks to several providers and to open-weight models you host. Swapping a model is a policy change, not a rewrite of your agents.
Where your policy says. Voice AI today runs on Google Cloud in Mumbai with open-weight models, designed for sub-second turns. See the deployment page for the options.
Yes. Policy is set per tenant, so each business line can have its own chain and its own limits.
Bring your own model policy.
Tell us which models your risk team allows. We will show the router working under that policy.