Show HN: I built and AI phone system and wrote a step by step instructions

Why AI Voicemail Systems Matter

In today’s fast-paced world, our phones are constantly busy with customer inquiries, urgent requests, and follow-ups. Traditionally, businesses relied on call centers for large enterprises, receptionists for mid-sized companies, and personal assistants for busy individuals, but AI-powered phone services now offer a reliable and cost-effective alternative. These systems can handle high call volumes, provide immediate customer service, and operate 24/7, ensuring no call goes unanswered even outside office hours or while traveling internationally.

Use Cases:

• AI front desk
• Customer support
• Sales qualification
• Automated voicemail with transcript and summary

High-Level Architecture Overview

The system consists of four main components:

1. Twilio Media Streams – Streams live call audio to your server.
2. FastAPI WebSocket Bridge – Connects Twilio ↔ OpenAI and handles audio conversion.
3. OpenAI Realtime API – Processes live AI conversation and generates responses.
4. Supabase – Stores call transcripts, AI summaries, and voicemail data.

Flow Diagram:

Key Components and Technologies

Prerequisites

1. Accounts & Services

   • Twilio account + phone number
   • OpenAI API key (Realtime API access)
   • Supabase project with tables: calls, call_transcripts, user_settings, knowledge_base

2. Python Environment

   • Python 3.8+
   • pip package manager

3. Development Tools

   • ngrok for local testing
   • Terminal/command line access

Step-by-Step Setup Instructions

Step 1: Install Dependencies

Step 2: Configure Environment Variables

Create a .env file:

Step 3: Set Up Supabase Tables

Create the following tables:

• calls: call metadata
• call_transcripts: transcripts and summaries
• user_settings: phone number → user_id
• agent_prompts: custom prompts for AI
• knowledge_base: optional RAG chunks

Step 4: Create the Twilio Webhook

• Endpoint: /api/v1/incoming-call-realtime
• Returns TwiML connecting call to WebSocket

Step 5: Write the WebSocket Bridge

• Accepts Twilio WebSocket
• Connects to OpenAI Realtime API
• Handles media events: connected, start, media, stop
• Collects transcripts and forwards audio

RAG schema

Audio Format Conversion

• Twilio → OpenAI: μ-law 8kHz → PCM16 24kHz
• OpenAI → Twilio: PCM16 24kHz → μ-law 8kHz
• Conversion handled by audioop-lts library

Streaming Conversation Logic

• Twilio sends media events → converted → OpenAI
• OpenAI sends response.audio.delta → converted → Twilio
• AI transcription events collected in real-time
• Async tasks handle bidirectional streaming

Saving Transcripts to Supabase

• Collect conversation lines (user + AI)
• Generate AI summary
• Save to calls and call_transcripts tables

Testing the System

• Start FastAPI: uvicorn app.main:app --reload --port 8000
• Expose with ngrok: ngrok http 8000
• Update Twilio webhook to ngrok URL
• Make test calls, verify audio streaming, transcription, and RAG-enhanced responses

What You Can Build With This System

• AI Receptionist – Answer calls automatically
• Customer Support Bot – Live issue resolution
• Sales Qualification Agent – Collect leads
• AI Voicemail System – Automated greeting, recording, transcript, summary
• Multi-Tenant SaaS – Custom AI agents per business
• Internal Helpdesk – HR, IT support
• Workflow Automation – Trigger notifications or CRM actions

TL;DR

Either read this article or feed IDE of your choice with this context and let it run it for you. Download ready-to-use prompt

By following this guide, you can launch build an AI voice agent capable of handling calls, transcribing them, and generating voicemail summaries automatically. Happy vibing!