Case studyLive

AI Receptionist

A real-time voice AI receptionist that answers calls, books appointments, captures waitlists, and escalates to a human — with quick-reply caching that hits sub-100ms response paths for high-frequency intents.

Role
Builder · Personal project
Timeline
2025
Links
Visit
Next.jsTypeScriptSupabaseRetell AIGPT-4o-miniGemini 2.0 Flash
<100ms
hot-path responses
24/7
call answering
14+
countries served
12
business verticals
01

The problem

Businesses lose real revenue to the phone. Front desks are busy with customers in the room, after-hours calls roll to voicemail, and every missed call is a missed booking — or a customer who calls the next business instead.

Hiring and training reception to cover that volume is expensive and doesn't scale. Meanwhile callers expect an immediate answer to simple things: opening hours, availability, booking, rescheduling.

An old-school phone tree ("press 1 for…") only frustrates people. The bar is a conversation that actually understands and acts.

02

The approach

An AI voice agent — not an IVR — that picks up, understands intent in natural speech, and takes action: answering FAQs, booking appointments, capturing waitlists, and routing urgent or complex cases to a human.

It's built on Retell AI for low-latency real-time voice, with GPT-4o-mini and Gemini 2.0 Flash driving LLM tool-calling into Supabase and the Google Calendar API. A quick-reply cache short-circuits high-frequency intents to hit sub-100ms response paths.

A Next.js + Supabase dashboard exposes call logs, scheduling, analytics, and billing — secured with PII encryption, Row-Level Security, and webhook-driven automations.

03

Architecture

Callrealtime
Inbound call
caller
Retell AI
STT · LLM · TTS
Quick-reply cache
<100ms hot paths
LLM tool-calling
GPT-4o-mini · Gemini
Action
book · waitlist · escalate
Spoken reply
via Retell
Platformconfig
Next.js dashboard
logs · scheduling · billing
Supabase
RLS · PII encryption
Google Calendar
appointment sync
Webhooks
automations

// Tool-calling writes through to Supabase and Google Calendar; urgent or complex calls are escalated to a human.

04

How it's built

  • Integrated Retell AI for natural, low-latency voice and wired its webhooks into LLM tool-calling with GPT-4o-mini and Gemini 2.0 Flash.
  • Built intent handling — appointment booking, waitlist capture, and human escalation — backed by Supabase and the Google Calendar API.
  • Added a quick-reply cache for high-frequency intents to hit sub-100ms response paths.
  • Shipped a Next.js + Supabase dashboard for call logs, scheduling, analytics, and billing — secured with PII encryption, Row-Level Security, and webhook-driven automations.
05

The outcome

Every call gets answered — 24/7 — so after-hours and overflow bookings stop leaking to voicemail or competitors.

Quick-reply caching keeps high-frequency intents under 100ms, so the conversation feels immediate rather than laggy.

Deployed across 12 business verticals and 14+ countries, with PII encryption and Row-Level Security guarding customer data.

Next case study
GoParkEasy Platform
Production SaaS parking platform

Want the deep dive?

Happy to walk through the design decisions and trade-offs.