🚗 Vehicle Insurance Call AI Agent

A sophisticated bilingual (Hindi/English) AI voice agent built with LiveKit that handles vehicle insurance inquiries, extracts customer information, and stores it in a PostgreSQL database. The agent uses a Neo4j Graph RAG system to retrieve relevant car information scraped from CarDekho.com using Firecrawl API.

🌟 Features

🎯 Core Capabilities

• Bilingual Voice Interaction: Seamless switching between Hindi and English
• Real-time Voice Conversations: Using LiveKit's advanced voice pipeline
• Lead Qualification: Intelligent conversation flow to assess customer interest
• Appointment Booking: Automated scheduling for qualified leads
• Data Extraction: Captures customer details, preferences, and vehicle information

🚀 Technical Features

• Graph RAG System: Neo4j-powered knowledge graph for intelligent car information retrieval
• Web Scraping: 200+ car models scraped from CarDekho.com using Firecrawl API
• Real-time Database: PostgreSQL with Prisma ORM for data persistence
• Cloud Infrastructure: Cloudflare Workers backend deployment
• Advanced Voice Processing: Noise cancellation, VAD, and multilingual support

🏗️ Architecture

graph TB
    A[LiveKit Voice Interface] --> B[AI Agent - Priya]
    B --> C[Sarvam STT/TTS]
    B --> D[Google Gemini LLM]
    B --> E[Neo4j Graph Database]
    B --> F[PostgreSQL Database]
    
    G[Firecrawl API] --> H[Car Data Scraper]
    H --> I[CSV Data Files]
    I --> E
    
    E --> J[Real-time Graph RAG Queries]
    J --> B
    
    F --> K[Cloudflare Workers API]
    K --> L[Prisma ORM]
    
    subgraph "Data Sources"
        M[CarDekho.com]
        M --> G
    end
    
    subgraph "Voice Stack"
        N[Silero VAD]
        O[Noise Cancellation]
        P[Turn Detection]
    end
    
    N --> A
    O --> A
    P --> A

Loading

📊 Data Pipeline

1. Car Data Collection

• Source: CarDekho.com (40+ manufacturers)
• Tool: Firecrawl API for structured extraction
• Coverage: 200+ models with complete specifications
• Output: CSV files with detailed car information

2. Graph Database Construction

• Database: Neo4j Graph Database
• Structure: Vehicle nodes connected to Manufacturer, BodyType, FuelType, and Feature nodes
• Relationships: MANUFACTURED_BY, HAS_BODY_TYPE, USES_FUEL, HAS_FEATURE
• Purpose: Real-time car information retrieval during conversations via Graph RAG

3. Voice Conversation Pipeline

User Voice → STT (Sarvam Saarika v2.5) → LLM (Gemini 2.5 Flash) → Graph RAG → TTS (Sarvam Anushka) → User
                                                    ↓
                                            PostgreSQL Storage

🤖 AI Models & Configuration

Voice Processing Models

• STT Model: Sarvam Saarika v2.5 (Hindi/English bilingual)
• TTS Model: Sarvam Anushka voice (Hindi/English bilingual)
• Language Model: Google Gemini 2.5 Flash
• Voice Activity Detection: Silero VAD
• Turn Detection: Multilingual model for seamless conversation flow

Model Configuration

stt=sarvam.STT(
    language="hi-IN",
    model="saarika:v2.5",
),
llm=google.LLM(model="gemini-2.5-flash"),
tts=sarvam.TTS(
    target_language_code="hi-IN",
    speaker="anushka",
),

🛠️ Tech Stack

AI & Voice Processing

• LiveKit Agents: Voice conversation framework
• Vapi Integration: Voice AI platform for enhanced conversation capabilities
• Sarvam STT: Speech-to-Text using Saarika v2.5 model (Hindi/English)
• Sarvam TTS: Text-to-Speech using Anushka voice (Hindi/English)
• Google Gemini 2.5 Flash: Large Language Model for conversation generation
• Silero VAD: Voice Activity Detection
• Multilingual Turn Detection: Conversation flow management

Backend Infrastructure

• Cloudflare Workers: Serverless backend deployment
• Hono.js: Fast web framework
• Prisma ORM: Database management
• PostgreSQL: Primary database
• TypeScript: Type-safe development

Data & AI

• Neo4j: Graph database for RAG system
• Firecrawl API: Web scraping for car data
• Pandas: Data processing and analysis
• LangChain: Graph database integration

Development Tools

• Python: Agent and scraping scripts
• Node.js: Backend services
• Jupyter Notebooks: Data analysis and graph setup

📁 Project Structure

Vehicle-Insurance-Agent/
├── agent.py                    # Main LiveKit agent
├── scrapper.py                 # Car data scraper
├── models.py                   # Pydantic data models
├── requirements.txt            # Python dependencies
├── env.local.example          # Environment variables template
│
├── Backend/                    # Cloudflare Workers API
│   ├── src/index.ts           # API entry point
│   ├── routes/user.ts         # User management endpoints
│   ├── prisma/
│   │   ├── schema.prisma      # Database schema
│   │   └── migrations/        # Database migrations
│   ├── package.json           # Node.js dependencies
│   └── tsconfig.json          # TypeScript configuration
│
├── data/                      # Scraped car datasets
│   ├── car_dataset_combined.csv    # Main dataset (200+ models)
│   └── car_data_*.csv              # Individual scraping runs
│
├── notebooks/                 # Data analysis and setup
│   ├── data.ipynb            # Data processing notebook
│   └── graphdb.ipynb         # Neo4j graph setup
│
└── public/                   # Graph visualizations
    ├── graph1.png           # Neo4j graph structure
    └── graph2.png           # Vehicle relationships

🚀 Setup Instructions

1. Clone Repository

git clone <repository-url>
cd Vehicle-Insurance-Agent

2. Environment Setup

# Copy environment template
cp env.local.example .env.local

# Install Python dependencies
pip install -r requirements.txt

# Install Node.js dependencies
cd Backend
npm install

3. Configure Environment Variables

Edit .env.local with your API keys:

# Google API Configuration
GOOGLE_API_KEY=your_google_api_key_here

# LiveKit Configuration
LIVEKIT_URL=wss://your-livekit-instance.livekit.cloud
LIVEKIT_API_KEY=your_livekit_api_key
LIVEKIT_API_SECRET=your_livekit_api_secret

# Sarvam API Configuration
SAKVAM_API_KEY=sk_your_sarvam_api_key

# FireCrawl API Configuration
FIRECRAWL_API_KEY=fc-your_firecrawl_api_key

# Neo4j Configuration
NEO4J_URI=neo4j+s://your-instance.databases.neo4j.io
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=your_neo4j_password

# Groq API Configuration (optional)
GROQ_API_KEY=gsk_your_groq_api_key

4. Database Setup

PostgreSQL (Prisma)

cd Backend
npx prisma generate
npx prisma db push

Neo4j Graph Database

# Run the graph setup notebook
jupyter notebook notebooks/graphdb.ipynb

5. Data Collection (Optional)

# Scrape fresh car data from CarDekho.com
python scrapper.py

6. Run Services

Start Backend API

cd Backend
npm run dev

Start AI Agent

python agent.py

🧠 Graph RAG in Action

Real-time Car Query Processing

The Neo4j Graph RAG enables:

• Instant Answers: Sub-second response to car specification queries
• Contextual Understanding: Links between manufacturers, models, features
• Complex Queries: Multi-attribute searches across 200+ car models
• Personalized Responses: Based on user's specific car variant

🎯 Agent Conversation Flow

1. Language Selection

Hindi: "नमस्ते! मैं प्रिया बोल रही हूँ SecureWheels Insurance से। 
       आप हिंदी में बात करना चाहेंगे या English में?"

English: "Hello! This is Priya from SecureWheels Insurance. 
         Would you prefer Hindi or English?"

2. Information Gathering

• Customer name and contact details
• Vehicle ownership information
• Specific car model details
• Current insurance status

3. Real-time Car Information Retrieval (Graph RAG)

Mid-conversation, whenever users ask about car details, the agent:

• Queries the Neo4j Graph Database in real-time
• Retrieves specific car specifications, features, and pricing
• Provides instant answers about mileage, engine specs, safety ratings
• Offers personalized insurance recommendations based on car data
• Handles complex queries like "What's the mileage of Tata Nexon XZ variant?"

Example Mid-Conversation:

User: "I have a Tata Nexon. What's its mileage?"
Agent: *Queries Neo4j Graph DB*
Agent: "The Tata Nexon offers excellent mileage! The petrol variant gives 17.57 kmpl 
        while the diesel variant delivers 21.99 kmpl. Which variant do you have?"

4. Lead Qualification & Booking

• Assess customer interest (1-10 score)
• Analyze sentiment (Positive/Neutral/Negative)
• Book appointments for qualified leads
• Store complete interaction data

📈 Dataset Statistics

• Manufacturers: 40+ brands (Tesla, Tata, Maruti, Hyundai, BMW, etc.)
• Total Models: 200+ different car models
• Data Points: 15+ attributes per vehicle
• Coverage: Complete specifications, pricing, features
• Source: CarDekho.com

🎨 Graph Visualizations

The project includes Neo4j graph visualizations showing:

1. Graph Structure (graph1.png): Overall node relationships and data model
2. Vehicle Connections (graph2.png): How vehicles connect to features, manufacturers, and specifications
3. Advanced Graph Analysis (graph3.png): Extended graph relationships and data insights