Real-time voice personas for insurance workflows
The hardest parts of voice AI are the live ones: sub-second turn-taking, interruption, keeping a persona stable, and staying in control while two audio streams run. We built a demonstration that does all of it in real time — the groundwork for phone workflows like insurance claims triage and quote qualification.
Laravel · Reverb WebSockets · Gemini Live · React
Ref 01 · The problem
The problem
- Insurance operations run on high-volume, structured phone conversations — first-notice-of-loss, claims triage, quote qualification — exactly the shape of call voice AI should handle.
- Most "voice AI" demos are turn-based request-response; a credible agent has to stream, interrupt and be interrupted like a person.
- Any live AI conversation needs an operator’s hand on the controls: mute, redirect, stop.
Ref 02 · What we built
What we built
Genuinely real-time conversation
Audio streams both ways over WebSockets to Gemini Live — streamed PCM, not request-response — so responses begin while the other side is still finishing, and interruptions land naturally.
Distinct, isolated voice personas
Each persona runs an isolated session with its own system prompt and its own voice — the isolation you need for an agent that represents your brand call after call.
A moderator console with real authority
Live controls — interrupt, force-speaker, mute, kill switch — broadcast over Laravel Reverb WebSockets, plus a safety-filter layer that counts and surfaces blocked content. The human stays in charge of the call.
Ref 03 · Stack
Stack
- Laravel
- Reverb WebSockets
- Gemini Live
- React
Ref 04 · What it demonstrates
What it demonstrates
- The full real-time stack: streaming audio, turn-taking orchestration, live control — the parts a slide deck can’t fake
- Persona engineering: isolated, constrained AI speakers rather than a generic assistant voice
- A direct path to claims-triage and quote-qualification agents, built on the same real-time voice core as our HGV screening system