Vilozi Engineering
· 7 min read
A missed call in the trades is usually a missed job. The person with a leaking roof calls the next number on the list. So the first thing our voice work had to get right wasn't the AI at all. It was making sure a call always reaches someone, person or agent, on the first ring.
One inbox, every channel
Calls, texts and website chat all land in one live inbox that every team member carries on their phone. A website visitor's message reaches the phone in about four seconds, and the reply shows up in their chat window while they're still on the page. The voice agent works from the same inbox, so a conversation can move between AI and people without anyone losing the thread.
Ringing a locked phone
The hardest engineering in the project was ordinary-sounding: making an incoming call ring a phone that is locked, in a pocket, with the app closed. Mobile operating systems are built to save battery, not to wake an app for an unexpected call. Getting a reliable ring meant push notifications tuned for calls, a calling kit patched to ring at full ringer volume until answered, declined or hung up, and tests on real hardware for every step.
Today a call rings through on a locked screen, can be answered and talked on, and stops ringing within about a second of the caller hanging up.
When the AI answers
When nobody on the team can pick up, the voice agent does. It is built to do a small number of things well: understand what the job is, ask the questions a dispatcher would ask, check the schedule, and book the visit. It doesn't pretend to be a person, and it hands off to a person when the caller asks or when the job is outside what it should handle.
Fast enough to feel natural
Voice is unforgiving about delay. A pause that would go unnoticed in a chat feels broken on the phone. We keep the agent fast by routing each step to the smallest model that does it well, and saving larger models for the moments that need them, like turning a long description of a job into a priced quote. More on that in One task, many models.
Vilozi Engineering
Notes from the Vilozi engineering team on the open problems we work on.
