Reliability

How the platform stays available during a provider fault, what is measured, and the targets it works to.

The platform's priority order is reliability first, then call quality, then latency.

Answering a call

Several checks run before a call is answered. Each one is designed so that its own failure cannot cost a customer the call.

Call arrives
  1. number routing   lookup error -> refuse the call
  2. billing check    error        -> answer the call anyway
  3. capacity check   at ceiling   -> answer, then transfer or apologise
  4. warm-up          over 2s      -> answer the call anyway

Number routing

The dialled number is resolved to an agent. If that lookup returns an error, the call is refused before it is answered rather than being routed to a default agent.

Billing check

Fails open. A billing error never prevents a call being answered.

Capacity check

If the platform is at its ceiling, the call is still answered:

With an overflow forward number set: the caller is told a human is being connected, and the call is transferred.
Without one: the caller hears a short apology and the call ends.

While the caller still hears ringing, the platform loads the agent settings and opens a connection to the speech synthesis provider, under a hard two-second budget. If that does not complete, the call is answered anyway.

The model's prompt cache warm and the pre-rendered greeting audio run in the background and land during the greeting rather than holding the caller.

Failover

Language model failover is active in production. Cross-provider voice failover is available for agents with a fallback voice configured.

Failover covers each subsystem in detail.

Self-healing

A stalled runtime is detected and restarted automatically. The check is written so that it can never act on a machine that is serving a call.

Deployment

Releases do not interrupt calls in progress. A new build must pass health checks before anything serving calls is touched. Machines running the previous build drain: they stop taking new calls and keep in-flight calls connected until they end.

Latency

The target is under 500 milliseconds from a caller finishing speaking to hearing any audio in response.

Per-turn latency is recorded for every call and displayed in the product. Latency depends on the model and voice you chose, so measure your own agent.

Backups and recovery

The database is managed by our hosting provider with standard backup and point-in-time recovery. Secret material is also mirrored to a separate encrypted store in the same Australian region.