Loan Approval Prediction System

Project Overview

This is an end-to-end Machine Learning system that predicts loan approval status based on user input. The project simulates a real-world deployment pipeline, integrating model serving, user interface, API gateway, and monitoring stack — all containerized with Docker and orchestrated via GitHub Actions.

✅ Test-deployment structure
🔁 CI/CD tested via GitHub Actions
📊 Real-time monitoring with Prometheus + Grafana

---

Problem Statement

Manual loan approval processes in financial institutions are often slow and inconsistent. This project demonstrates how machine learning can automate and standardize such decisions, backed by modern MLOps practices.

The goal is to:

• Automate loan approval decision-making
• Serve a model through a REST API
• Provide a friendly UI for interaction
• Monitor system performance and behavior

---

Dataset

The raw dataset version was downloaded from this link. This project uses a preprocessed dataset of loan applicants, including features like:

• Age
• Income
• Loan amount
• Loan interest
• Loan percent income
• Credit score
• Previous loan default indicator

The target variable is a binary label indicating loan approval (1 = approved, 0 = rejected).

---

Tech Stack

Layer	Tools
Model Training & Serving	MLflow
API Gateway	Flask
User Interface	Streamlit
Monitoring	Prometheus + Grafana
Containerization	Docker
CI/CD	GitHub Actions (Docker build + deploy)

---

How It Works

Components:

1. ML Model (mlflow)

   • Utilize dagshub to store experiments and model artifacts
   • Registered and loaded using MLflow’s model registry

2. API (Flask)

   • Receives prediction requests from the UI
   • Forwards data to the ML model container
   • Returns prediction + probabilities
   • Logs metrics to Prometheus

3. UI (Streamlit)

   • Allows users to input loan applicant data
   • Connects to the Flask API
   • Displays the prediction results

4. Monitoring

   • Prometheus scrapes metrics from the API
   • Grafana visualizes metrics like:
     • API latency
     • Request count
     • CPU/RAM usage
     • Model confidence distribution

---

CI/CD & Test Deployment

Deployment is handled using GitHub Actions:

• Builds & pushes:
  • loan-streamlit
  • loan-api
  • loan-mlmodel
• Deploys & tests:
  • Launches all containers via docker-compose
  • Waits for streamlit, api, prometheus, and grafana to pass health checks
  • Prints logs for debugging if services fail

Located in .github/workflows/deploy-compose.yml

---

Monitoring Stack

• API exposes Prometheus metrics at /metrics
• Preconfigured prometheus.yml sets up scraping
• Grafana is provisioned with Prometheus as a data source

---

Project Setup (Local Development)

Ensure Docker, Prometheus, and Grafana are installed.

1. Clone the Repository

git clone https://github.com/yourusername/loan-approval-predictor.git
cd loan-approval-predictor

2. Set up virtual environment

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install Dependency

pip install -r requirements.txt

4. Set up environment variable using dotenv

Create .env file, then fill in the necessary keys

MLFLOW_TRACKING_USERNAME=your_mlflow_username
MLFLOW_TRACKING_PASSWORD=your_mlflow_tracking_password
MLFLOW_TRACKING_URI=your_mlflow_tracking_uri
MLFLOW_EXPERIMENT_NAME=your_mlflow_experiment_name
MLFLOW_RUN_ID=your_model_run_id
DOCKER_HUB_USERNAME=your_docker_hub_username

5. Run all services

docker compose up --build

6. Access services

Service	URL
Streamlit UI	http://localhost:8501
Flask API	http://localhost:8000/health
Prometheus	http://localhost:9090
Grafana	http://localhost:3000

---

Limitation

• This project is not ready for live predictions; it's still in the test-deployment phase.
• While CI/CD pipelines for Docker images (model, API, UI, Prometheus, Grafana) are built and run with Docker Compose, I haven't done local tests due to a Docker Compose issue on my device. End-to-end testing to confirm all services run smoothly is highly needed.

Future Improvements

• End-to-end testing
• Add testing (unit + integration)
• Enhance UI usability and model explainability
• Add service-level authentication (Flask, Grafana)
• Push to production deployment (GCP, ECS, Azure, etc.)

---

Author

Built by Roissyah Fernanda, this is part of a broader effort to create full-stack machine learning pipelines with real deployment readiness. Feel free to send pull request, Let's collaborate!.