Fraud Job Detection

An end-to-end Machine Learning application that classifies job postings as Fraudulent or Legitimate using NLP and structured features. The project includes a trained ML pipeline, a FastAPI prediction service, and a Streamlit demo UI for real-time inference.

Status: End-to-end prototype (model + API + frontend)

---

System Architecture

---

📌 Project Highlights

• 🧠 Trained ML pipeline (scikit-learn / XGBoost)
• 🚀 FastAPI backend serving predictions
• 🎛 Streamlit UI for easy testing
• 📦 Model artifacts versioned and reusable
• 🐳 Docker-ready backend (optional deployment)

---

📂 Repository Structure

• app/app.py — FastAPI app exposing / (health) and /predict
• app/predict.py — Model loading & prediction logic
• st.py — Streamlit demo UI (calls API)
• model/ — Saved model artifacts
  • fraud_detection_pipeline.pkl
  • decision_threshold.pkl
• data/ — CSV datasets used during development
• model_notebooks/ — Training & evaluation notebooks
• data_cleaning/, data_processing/ — Preprocessing notebooks
• Dockerfile — Backend container configuration
• requirements.txt — Python dependencies (Backend)
• requirements2.txt — Python dependencies (Frontend)

---

🚀 Getting Started (Local)

1️⃣ Setup environment

python -m venv .venv
.\.venv\Scripts\activate
pip install -r requirements.txt

2️⃣ Run the API

uvicorn app.app:app --host 0.0.0.0 --port 8000 --reload

Health check:

GET http://localhost:8000/

Response:

{"status":"API is running"}

---

🔮 Prediction Example

curl -X POST "http://localhost:8000/predict" \
  -H "Content-Type: application/json" \
  -d '{
    "job_description":"Work from home opportunity...",
    "telecommuting":1,
    "has_company_logo":0,
    "has_questions":0,
    "jd_word_count":18,
    "req_exp_enc":0,
    "employment_type_Full_time":0,
    "employment_type_Part_time":0,
    "employment_type_Temporary":0,
    "employment_type_Other":1,
    "employment_type_Unknown":0
  }'

Response includes:

• fraud_probability
• prediction (Fraud / Real)
• threshold_used

---

🧪 API Schema Notes

• The API expects the same feature names used during training
• Input combines:
  • Job description text
  • Binary metadata features
  • One-hot encoded employment types
• See app/app.py for the exact schema

---

🎛 Streamlit Demo

1. Update API URL inside st.py:

API_URL = "http://localhost:8000/predict"

2. Run:

streamlit run st.py

The Streamlit app includes retry logic to handle cold-start delays on free hosting platforms.

---

🐳 Docker (Optional)

The backend is Docker-ready and can be containerized later without code changes:

docker build -t fraud-job-detection .
docker run -p 8000:8000 fraud-job-detection

---

📈 Model Artifacts

Stored in model/ and loaded at runtime:

• fraud_detection_pipeline.pkl — preprocessing + classifier
• decision_threshold.pkl — probability threshold

---

🔧 Future Improvements

• Add CI pipeline for linting & API smoke tests
• Add unit tests for prediction logic
• Docker Compose for API + Streamlit
• Model monitoring & logging

---

📜 License

MIT License

---

👤 Author

Rajdip

If you found this project useful, feel free to ⭐ the repository.