work / callconvert · client work · 2025–2026
CallConvert
An AI receptionist that answers, books, and follows up — 24/7
system
caller ──▸ AI voice agent ──▸ emergency? ──▸ escalate
│ to human
├─ books appointment ─▸ calendar
├─ answers from RAG over the
│ business's own documents
└─ missed lead ─▸ auto-chase workflow
│
visual workflow editor (the client
configures all of this themselves)the job
Home-service businesses lose work to unanswered phones. The product answers every call with an AI voice agent that can detect an emergency, book an appointment against real availability, answer questions from the business's own documents, and chase leads that didn't convert.
what I built
The full platform: voice-agent configuration (script, voice, vocabulary, escalation handling), a visual workflow editor so non-technical owners can change the automation themselves, retrieval over uploaded business docs with pgvector, a BullMQ worker fleet for async jobs, and the commercial machinery — Stripe billing, onboarding funnel, admin console.
The hard part of voice AI isn't the demo, it's the edges: interruptions, bad audio, callers who change their mind mid-booking, and making an agent hand off to a human the moment it should. Most of the codebase exists to handle exactly those.
scale
This is a real production system for a client — around 83,000 lines across a Turborepo monorepo, 270+ commits, still shipping weekly.
stack
Next.js · Turborepo + Bun · Supabase + pgvector · Vapi · BullMQ · Twilio · Stripe
Client engagement — codebase private.