Vehicle Insurance Cross-Sell Predictor

A production-grade MLOps pipeline that predicts whether a health insurance customer is likely to purchase vehicle insurance - built end-to-end from data ingestion to live cloud deployment.

Live Demo: http://3.87.31.9:8080

The Problem

Insurance companies have millions of existing health insurance customers. Reaching out to all of them to sell vehicle insurance is expensive and inefficient. This project builds a machine learning model that identifies which customers are actually likely to say yes, so the sales team can focus their efforts where it counts.

What Makes This MLOps

This isn't just a Jupyter notebook. Every component mirrors how ML systems are built and maintained in production:

• Automated data pipeline - pulls fresh data from MongoDB, validates schema and detects data drift before any training happens
• Model evaluation gate - new models only go to production if they beat the existing model by a measurable threshold
• Cloud model registry - trained models are versioned and stored in AWS S3, not on someone's laptop
• Zero-downtime deployments - every push to GitHub automatically builds, containerizes, and deploys the updated app to EC2 without any manual steps
• Reproducible environments - Docker ensures the app runs identically in development and production

Tech Stack

Layer	Technology
Data Storage	MongoDB Atlas
Model Training	Scikit-learn, XGBoost
Model Registry	AWS S3
API	FastAPI
Containerization	Docker
Container Registry	AWS ECR
Cloud Server	AWS EC2
CI/CD	GitHub Actions

ML Pipeline

flowchart LR
    A[(MongoDB Atlas)] --> B[Data Ingestion]
    B --> C[Data Validation]
    C --> D[Data Transformation]
    D --> E[Model Training]
    E --> F{Evaluation Gate}
    F -->|beats production| G[(AWS S3\nModel Registry)]
    F -->|below threshold| H[Keep Production Model]
    G --> I[Model Pusher]

Loading

CI/CD Pipeline

flowchart LR
    A[git push] --> B[GitHub Actions]
    B --> C[CI Check]
    C --> D[Build Docker Image]
    D --> E[Push to AWS ECR]
    E --> F[EC2 Self-hosted Runner]
    F --> G[Pull Latest Image]
    G --> H[Restart Container]
    H --> I[Live at 3.87.31.9:8080]

Loading

Every code change is automatically tested, built, and deployed. No manual SSH. No manual restarts.

Architecture

flowchart TB
    subgraph Data["Data Layer"]
        A[(MongoDB Atlas)]
    end

    subgraph Training["Training Pipeline"]
        B[Ingestion] --> C[Validation]
        C --> D[Transformation]
        D --> E[Training]
        E --> F[Evaluation]
    end

    subgraph Registry["Model Registry"]
        G[(AWS S3)]
    end

    subgraph Serving["Serving Layer"]
        H[FastAPI App]
        I[Docker Container]
    end

    subgraph Cloud["AWS Cloud"]
        J[ECR]
        K[EC2]
    end

    subgraph CICD["CI/CD"]
        L[GitHub Actions]
    end

    A --> B
    F -->|accepted| G
    G --> H
    H --> I
    I --> J
    J --> K
    L -->|auto deploy| K

Loading

Some Decisions Worth Explaining

XGBoost over other models - The dataset is heavily imbalanced, only about 12% of customers actually respond yes. I tried logistic regression and random forest during experimentation but XGBoost consistently gave better F1 scores on the minority class without needing much manual tuning.

The evaluation gate - The naive approach is to just always deploy the newest model. But what happens when training runs on bad data, or someone tweaks the preprocessing and breaks something? The gate compares the new model against whatever is currently in S3 and only swaps it in if the F1 improves by at least 2%. It's a small safety net that prevents silent model degradation.

Running Locally

You'll need Python 3.10+, a MongoDB Atlas account, and AWS credentials with S3 access.

Clone the repo and install dependencies:

git clone https://github.com/VinaySampath14/vehicle_insurance_mlops.git
cd vehicle_insurance_mlops
pip install -r requirements.txt

Create a .env file in the root with:

MONGODB_URL=your_mongodb_connection_string
AWS_ACCESS_KEY_ID=your_key
AWS_SECRET_ACCESS_KEY=your_secret

Run the app:

python app.py

Hit /train once to train the model and push it to S3. After that /predict works.

API Endpoints

Endpoint	Method	Description
/	GET	Prediction form UI
/predict	POST	Returns 0 or 1 for a given customer profile
/train	GET	Triggers the full training pipeline
/docs	GET	Interactive Swagger API documentation

Dataset

The dataset contains ~380,000 records of existing health insurance customers with features including age, vehicle age, prior damage history, annual premium, and sales channel. The target variable (Response) indicates whether the customer expressed interest in vehicle insurance.

Author

Vinay Sampath Kumar Vudumula GitHub