Reliability
The AI wave moves fast. Your phone line shouldn't.
A missed call is a missed booking, a frustrated member, a lost customer. So we engineer the boring half of voice AI as seriously as the clever half. This page shows how, in detail.
The numbers below come from failover drills on live production traffic, not from a lab.
Architecture
Built to lose a datacenter
Every serious outage story starts with a single point of failure. We removed ours: the platform runs on two identical stacks in separate datacenters, with one floating address in front.
One floating address
Carriers and SIP trunks point at a single address that can move between datacenters in seconds. Your number, your trunk and your configuration never change during a failover.
A warm standby, not a cold spare
The second datacenter runs the full stack continuously: voice agents registered, telephony listening, certificates in place. It refreshes its configuration from the primary every day.
A watchdog that can't fail silently
An independent watchdog checks the primary every 15 seconds and moves traffic after repeated confirmed failures. The watchdog itself is guarded by a dead-man's switch: if it stops reporting, an engineer gets called.
Rehearsed on production
We don't trust designs, we trust drills. We have deliberately failed the primary during live traffic and watched real calls get answered by the standby, with two-way audio, before failing back.
Release discipline
Stability without staleness
Fast-moving platforms break things weekly. Frozen platforms fall behind the AI wave. We refuse both: new models land within days, but every change travels through the same gate before it may touch production.
$ failover-watchdog --tail
[check] primary · https 200 · sip reachable (every 15s)
[hb] normal-state ping → uptime monitor (every 60s)
[drill] primary made unreachable, on purpose
[strike] primary DOWN 1/4 · world reachable ✓
[strike] primary DOWN 4/4 · outage confirmed
[flip] floating address → standby · 9.4s
[verify] inbound call answered · two-way audio ✓
✓ drill passed · failed back · zero client changesMonitoring
We know before you do
Every layer reports its health continuously: heartbeats every minute, synthetic checks from multiple regions, and paging that wakes an engineer. Below is what our failover drill actually looked like from the watchdog's perspective.
Talk to the team that runs it
If your phone line carries bookings, members or patients, stability is a requirement, not a feature request. We are happy to walk you through the architecture in detail.