High-Level Architecture for Your Medicare AI Bot + Vicidial

VICIdial → SIP Call → AI Voice Bot (TTS/STT + GPT) → Response Flow
                                 ↓
         CRM + Medicare API ↔ AI Bot ↔ Agent Assist or Transfer

Step-by-Step Integration Plan

1. Choose a Platform Setup

Let’s go with Option 2 (Semi-Custom) for control + Medicare compliance.

2. Core Components You’ll Need

3. VICIdial Configuration

Add SIP Carrier (e.g., Twilio)


[freeswitch]
type=friend
host=sip.freeswitch.com
username=YOUR_USER
secret=YOUR_SECRET
disallow=all
allow=ulaw
context=trunkinbound

Create a DID for the bot
- Point inbound calls to the SIP bot or route outbound auto-dials to it
Use a Remote Agent or In-Group Transfer
- Automatically route calls to the bot via campaign configuration

4. AI Voice Bot Application Logic

Build your AI voice bot handler in Node.js or Python, and connect it to Twilio or your SIP system:

Call Flow Example:

“Hello! This is your virtual Medicare assistant. Are you looking for Part C or D coverage?”
Transcribe user speech using Whisper/Deepgram
Process text using GPT-4 or LangChain prompt
Text-to-speech (ElevenLabs): “Great! Based on your ZIP code, you may qualify for zero-premium plans. Would you like me to connect you to a licensed agent?”

GPT Prompt Sample:


You are a friendly virtual assistant for Medicare plan comparisons.
Your goal is to:
- Ask basic eligibility (age, zip)
- Explain differences in Part C vs Part D
- Compare plans from a provided list (fetched via API)
- Help enroll or transfer to human

Always comply with CMS guidelines and speak slowly.

5. Compliance & Legal Notes

🔒 Must include:
- Recording consent
- CMS disclaimers (e.g., “We do not offer every plan available...”)
- Option to speak to a human
📜 Keep audit logs of each conversation and response
🎧 Store recordings via VICIdial or S3 bucket

6. Optional CRM & Reporting Layer

Log calls and transcripts to:
- Google Sheets (for prototype)
- Airtable
- Custom Laravel or Firebase backend
Use tags like:
- Interested, Needs Human, Unqualified, Do Not Call

Want an Example Stack?

🚀 Bonus: Outbound Campaign Strategy

Load leads into VICIdial with consent
Run a preview or auto-dial campaign
Send call to SIP AI Bot
Record + track responses
Route to live agent only if interested

Overview: System Architecture

plaintext
Copy

Edit[ VICIdial ]
    ↓ SIP Call
[ Twilio SIP / FS Gateway ]
    ↓ Webhook
[ AI Voice App (Node.js) ]
    ↓
[ STT (Whisper/Deepgram) ] → [ GPT-4 for NLP ] → [ TTS (ElevenLabs/Google) ]
    ↓
[ Speech Response → SIP Audio ]
    ↓
User

🧪 PoC Goals

GoalTechHandle outbound VICIdial call | VICIdial + SIP trunk
Route call to bot via SIP | Twilio Programmable Voice
Transcribe audio | Whisper or Deepgram
Use GPT to respond to Medicare queries | GPT-4 via OpenAI
Convert GPT text to speech | ElevenLabs TTS
Store call logs and transcripts | Firebase / MySQL
Optional: Handoff to human | Transfer via SIP or VICIdial campaign

🔧 Components Setup

1. VICIdial → SIP Setup

✅ Add Twilio SIP Trunk in VICIdial

ini
Copy

Edit[twilio]
type=peer
host=your-twilio-sip.sip.twilio.com
username=twiliouser
secret=twilioPassword
context=trunkinbound
disallow=all
allow=ulaw

✅ Create DID or Campaign

Route to this carrier
For outbound, use External Phones or Remote Agent

2. AI Voice Bot (Node.js)

Folder Structure

bash
Copy

Editai-voice-bot/
├── server.js
├── routes/
│   └── voice.js
├── services/
│   ├── stt.js
│   ├── gpt.js
│   └── tts.js
├── utils/
│   └── recorder.js
├── .env

server.js (Express App)

js
Copy

Editconst express = require('express');
const voiceRouter = require('./routes/voice');
const bodyParser = require('body-parser');
require('dotenv').config();

const app = express();
app.use(bodyParser.urlencoded({ extended: false }));
app.use('/voice', voiceRouter);

app.listen(3000, () => {
  console.log('Bot running on port 3000');
});

routes/voice.js

js
Copy

Editconst express = require('express');
const router = express.Router();
const { transcribe } = require('../services/stt');
const { generateReply } = require('../services/gpt');
const { synthesizeSpeech } = require('../services/tts');

router.post('/', async (req, res) => {
  const userSpeechUrl = req.body.RecordingUrl;
  const transcription = await transcribe(userSpeechUrl);
  const gptResponse = await generateReply(transcription);
  const audioUrl = await synthesizeSpeech(gptResponse);

  res.set('Content-Type', 'text/xml');
  res.send(`
    <Response>
      <Play>${audioUrl}</Play>
      <Record maxLength="30" action="/voice" method="POST" />
    </Response>
  `);
});

module.exports = router;

3. STT (Deepgram or Whisper)

services/stt.js

js
Copy

Editconst axios = require('axios');

async function transcribe(audioUrl) {
  const response = await axios.post('https://api.deepgram.com/v1/listen', {
    url: audioUrl
  }, {
    headers: { Authorization: `Token ${process.env.DEEPGRAM_API_KEY}` }
  });
  return response.data.results.channels[0].alternatives[0].transcript;
}
module.exports = { transcribe };

4. GPT Integration

services/gpt.js

js
Copy

Editconst { OpenAI } = require('openai');
const openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });

async function generateReply(inputText) {
  const response = await openai.chat.completions.create({
    model: "gpt-4",
    messages: [
      { role: "system", content: "You're a Medicare AI Assistant..." },
      { role: "user", content: inputText }
    ]
  });
  return response.choices[0].message.content;
}
module.exports = { generateReply };

5. TTS (ElevenLabs)

services/tts.js

js
Copy

Editconst axios = require('axios');
const fs = require('fs');
const path = require('path');

async function synthesizeSpeech(text) {
  const response = await axios.post(`https://api.elevenlabs.io/v1/text-to-speech/${process.env.ELEVENLABS_VOICE_ID}`, {
    text,
    voice_settings: { stability: 0.75, similarity_boost: 0.75 }
  }, {
    headers: {
      'xi-api-key': process.env.ELEVENLABS_API_KEY,
      'Content-Type': 'application/json'
    },
    responseType: 'arraybuffer'
  });

  const outputPath = path.join(__dirname, '..', 'public', `response.mp3`);
  fs.writeFileSync(outputPath, response.data);
  return `${process.env.PUBLIC_URL}/response.mp3`;
}

module.exports = { synthesizeSpeech };

📥 .env (sample)

ini
Copy

EditOPENAI_API_KEY=sk-...
DEEPGRAM_API_KEY=...
ELEVENLABS_API_KEY=...
ELEVENLABS_VOICE_ID=YOUR_VOICE_ID
PUBLIC_URL=https://your-domain.com

📡 Twilio Webhook

Set your SIP domain to hit:

arduino
Copy

Edithttps://your-domain.com/voice

✅ Deployment

Host on:
- Fly.io
- Vercel (with Express adapter)
- EC2 (if SIP routing locally)
Expose via HTTPS (Twilio requires it)

Fermin Perdomo