ConvoAI
A multi-persona voice AI receptionist with a protection system built from scratch.

Businesses lose calls they cannot answer. A voice receptionist has to feel instant and human, work across very different industries, and fail gracefully when speech models hiccup - or it erodes trust on the first call.
A voice AI receptionist with six industry personas, each with its own prompts, greetings, voices and theme. I built the entire core application, including a voice protection system that keeps the experience stable when the underlying speech providers fail.
VAPI Web SDK + Deepgram (Aura 2 voice, Flux transcriber)
Circuit breaker (3 errors / 60s), queue watchdog, start-call guards, provider fallback
Next.js + TypeScript, Better-Auth, PostgreSQL, server-side route protection
Real-time transcript + AI-actions feed, WebGL-style audio visualizer
Six personas
Healthcare, Real Estate, Sales, HR, Restaurant and Financial Services - each fully themed.
Voice protection
Error-rate circuit breaker with auto-recovery and a 60s stuck-call watchdog.
Live transcript
Real-time transcript and an AI-actions feed during the call.
Audio visualizer
A WebGL-style waveform that reacts to the live audio.
| Metric | Value |
|---|---|
| industry personas | 0 |
| circuit-breaker threshold | 0/60s |
| of the core app, solo | 0% |
| in production | Live |
I was the sole developer of the core application: the call interface, the multi-persona system with per-persona overrides, the real-time transcript and actions feed, the audio visualizer, and authentication. The voice protection system - circuit breaker, watchdog, start-call guards and provider fallback UX - was designed and built from scratch. I then handed off cleanly to a developer who extended it to real telephony.
- Real-time voice lives or dies on failure modes - the protection system mattered more than the happy path.
- Per-persona overrides kept one codebase serving six very different industries.
