CFLO AI Voice Agent

AI-powered outbound voice agent for debt collection and EMI recovery — built for CFLO, a fintech lending and collections platform.

---

What it does

The agent autonomously calls borrowers, identifies them by phone number, discusses overdue EMIs, handles hardship cases empathetically, and captures Promise-To-Pay (PTP) commitments — all without human intervention.

---

Architecture

Twilio (telephony) → FastAPI (backend) → Groq Llama 3.3 70B (LLM) → TwiML (voice response)
                            ↕
                     SQLite (mock banking DB)
                     PTP log / Call log / Borrower profiles

---

Tech Stack

Layer	Technology
Telephony	Twilio Voice
Backend	FastAPI + Uvicorn
LLM Brain	Groq — Llama 3.3 70B
Speech Recognition	Twilio built-in (Deepgram Nova-3 ready)
Text to Speech	Twilio Polly Aditi (ElevenLabs ready)
Database	SQLite (PostgreSQL ready for production)
Tunnel	ngrok
Language	Python 3.11

---

Features

• Outbound calling — agent calls borrowers automatically from dashboard
• Borrower identification — looks up caller by phone number in real time
• Intelligent dialogue — context-aware conversations using Llama 3.3 70B
• Tone calibration — empathetic for hardship cases, firm for chronic defaulters
• PTP capture — automatically logs payment commitments with amount and date
• Risk segmentation — S1 (critical) to S5 (self-cure) aligned with CFLO spec
• Compliance guardrails — no threats, no PII leakage, RBI Fair Practice Code aligned
• Management dashboard — view all borrowers, trigger calls, monitor PTPs
• Human escalation — hands off to live agent when needed

---

Project Structure

cflo-voice-agent/
├── main.py                    FastAPI app entry point
├── app/
│   ├── routes/
│   │   ├── voice.py           Twilio webhook handlers (inbound + outbound)
│   │   ├── outbound.py        Trigger outbound calls
│   │   └── dashboard.py       Management dashboard UI
│   └── services/
│       └── dialogue.py        LLM brain — Groq + system prompt + PTP logic
├── data/
│   └── database.py            SQLite — borrowers, PTP log, call log
├── .env                       API keys (not committed)
└── requirements.txt           Dependencies

---

Setup Instructions

1. Clone and install

git clone https://github.com/techAsmita/CFLO-Agent.git
cd CFLO-Agent
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

2. Create your .env file

TWILIO_ACCOUNT_SID=your_twilio_sid
TWILIO_AUTH_TOKEN=your_twilio_token
TWILIO_PHONE_NUMBER=your_twilio_number
GROQ_API_KEY=your_groq_key
BASE_URL=your_ngrok_url

3. Initialise the database

python data/database.py

4. Start the server

python main.py

5. Start ngrok tunnel

ngrok http 8000

6. Configure Twilio webhook

In Twilio Console, go to your phone number and set Voice Configuration webhook to:

https://your-ngrok-url/voice/inbound
HTTP Method: POST

7. Open the dashboard

http://localhost:8000/dashboard

---

Alignment with CFLO AI Specification

This prototype directly implements Section 4.4 (AI Voice Agent) of the CFLO AI-ML Architecture specification (v1.0, May 2026):

• Inbound and outbound call handling
• LLM core with intent classification and slot filling
• Dialogue state management with fallback paths
• PTP capture and logging
• Risk segment awareness (S1 critical through S5 self-cure)
• Hardship detection and empathetic tone switching
• Human escalation triggers
• Compliance guardrails aligned with RBI Fair Practice Code
• Deepgram Nova-3 STT integration ready
• ElevenLabs TTS integration ready
• Multi-language Hindi and Indic support ready

---

API Endpoints

Method	Endpoint	Description
GET	/	Health check
GET	/health	Server status
GET	/dashboard	Management dashboard
POST	/voice/inbound	Twilio inbound call webhook
POST	/voice/respond	Conversation turn handler
POST	/voice/outbound	Outbound call webhook
POST	/call/test	Trigger a test call

---

👩‍💻 Built by

Asmita — @techAsmita
B.E. Computer Engineering | AI/ML Enthusiast
LinkedIn