🛡️ Fraud Risk Intelligence System

Hybrid ML + GenAI fraud detection platform — XGBoost classifier with SHAP explainability and Anthropic Claude AI investigation reports.

Features

Feature	Detail
Fraud detection	XGBoost + SMOTE, trained on 50,000 transactions (1:578 imbalance handled)
Explainability	SHAP TreeExplainer — waterfall plot + top-5 risk factors per transaction
AI reports	Gemini (optional) writes investigation reports grounded in SHAP values
Bulk scoring	Upload any CSV, score all rows, download results with risk scores
Performance dashboard	ROC curve, PR curve, confusion matrix, feature importance

Quick start — Streamlit Cloud (free, recommended)

1. Fork this repo on GitHub
2. Go to share.streamlit.io → New app
3. Select your fork, branch main, file app.py
4. In Advanced settings → Secrets, paste:

   GEMINI_API_KEY = "sk-ant-your-key-here"

5. Click Deploy — live in ~2 minutes

Local setup

git clone https://github.com/SAMBIT-318/fraud-risk-intelligence.git
cd fraud-risk-intelligence

pip install -r requirements.txt

# Optional: add Gemini API key
cp .streamlit/secrets.toml.example .streamlit/secrets.toml
# Edit secrets.toml and replace the placeholder key

streamlit run app.py

Docker

# With Gemini (set your key in .env or export GEMINI_API_KEY=sk-ant-...)
docker-compose up --build

# App available at http://localhost:8501

Using the real Kaggle dataset (optional)

The app generates 50,000 synthetic transactions by default. For the full 284,807-transaction dataset:

1. Download creditcard.csv from Kaggle ULB Fraud Detection
2. Place it in the data/ folder
3. Restart the app — it auto-detects the real dataset

Architecture

Raw transaction (Amount, Time, V1–V28)
         ↓
Feature engineering  →  Amount_log, Amount_zscore, Hour, Is_night  (32 features)
         ↓
StandardScaler  →  SMOTE (train only)  →  XGBoost (200 estimators)
         ↓
Risk score (0–100)  +  SHAP TreeExplainer  →  top-5 feature attributions
         ↓
Gemini Google-genai gemini 2.5 flash  →  structured investigation report
         ↓
Streamlit (3 tabs): Analyzer | Bulk Scoring | Performance Dashboard

Risk thresholds

Score	Verdict
0–30	Legitimate
31–60	Needs review
61–85	High risk
86–100	Confirmed fraud

Skills demonstrated

Python · scikit-learn · XGBoost · SMOTE / imbalanced-learn · SHAP · Gemini API · Streamlit · Docker · Feature Engineering · Predictive Modeling · Plotly · Pandas · NumPy

Author

Sambit Swain — MCA Candidate | Data Science & AI/GenAI

GitHub · LinkedIn