FinSecure AI
Fraud signals with an explanation behind the score.
A fraud-detection project combining supervised classification, anomaly detection, and reconstruction error with a FastAPI service and an analysis dashboard.
The problem
An unusual transaction can be difficult to interpret through one model or a single opaque risk score, especially when fraudulent examples are rare.
How it comes together
The project trains on the Kaggle credit-card fraud dataset and combines XGBoost, Isolation Forest, and a PyTorch autoencoder into a weighted score. SHAP exposes contributing features.
Follow the flow.
Three-model ensemble with configurable alert thresholds.
Single and batch inference endpoints.
SHAP feature contributions and model-level insights.
Prediction logging, health checks, and dashboard analytics.
Why this approach?
Combine supervised and unsupervised signals, then expose explanations in the same workflow as inference so a risk score can be investigated.
A dataset-based engineering project; model performance depends on the evaluation data and deployment context.