Back
Year
2025
Tech & Technique
Python, FastAPI, TensorFlow, XGBoost, LSTM Autoencoder, Pandas, NumPy, PostgreSQL, Claude Sonnet 4, Bayesian Optimization, SMOTE, PCA, Docker
Description
CardFraudDetectAI is an AI-driven fraud detection system designed to identify both known fraud patterns and emerging hidden anomaly behaviors in real-world transaction streams. The goal was to reduce false fraud alerts while improving the detection of subtle behavioral shifts that traditional rule-based models miss.
Key Features:
Technical Highlights:
Key Features:
- 🚨 Dual-Model Detection Pipeline: Combines XGBoost (structured fraud signals) + LSTM Autoencoder (temporal anomaly detection)
- 📊 Feature Engineering System: Dynamic scaling, PCA reduction, and behavior sequence vectorization
- 🏦 Financial-Grade Risk Scoring: Produces interpretable fraud confidence scores per transaction
- ⚡ Low-Latency Prediction API: Built with FastAPI for real-time scoring workflows
- 🧠 Explainable AI: Claude Sonnet 4 generates human-readable reasoning summaries for flagged anomalies
Technical Highlights:
- 95% Precision on 1M+ Transactions: Achieved high precision and reduced false negatives through model ensembling and hyperparameter tuning
- LSTM Autoencoder for Behavioral Anomaly Detection: Captures temporal spending sequences and reconstructs deviation signatures to detect non-obvious fraud cases
- Class Rebalancing with SMOTE: Addressed severe class imbalance common in fraud datasets → boosted recall by 28%
- Bayesian Hyperparameter Optimization: Reduced false negatives by 22% while improving stability under different transaction distributions
- Human-Interpretable Fraud Reasoning: Leveraged Claude Sonnet 4 to convert cluster-level anomalies into explainable risk narratives, enabling audit compliance
My Role
AI Engineer • Data Scientist • System Architect
Research & Design:
Research & Design:
- Evaluated fraud detection models, selected hybrid XGBoost + LSTM architecture
- Built ETL pipelines, feature normalization flows, and temporal sequence datasets
- Performed SMOTE balancing, PCA reduction, Bayesian tuning, cross-model validation
- Deployed scoring API using FastAPI + Docker with async inference batching
- Integrated Claude to generate case summaries for analyst workflow integration
Case Study Impact
Problem: Traditional fraud systems over-rely on static rules, causing high false positives and missing new fraud behaviors.
Solution: Blend structural fraud pattern detection (XGBoost) with temporal anomaly detection (LSTM Autoencoder) and layer explainability to make signals actionable.
Outcome:
Solution: Blend structural fraud pattern detection (XGBoost) with temporal anomaly detection (LSTM Autoencoder) and layer explainability to make signals actionable.
Outcome:
- ✅ 95% precision on large-scale dataset
- ✅ 22% reduction in false negatives
- ✅ Detects emerging fraud patterns missed by baseline models
- ✅ Produces audit-ready reason codes for analysts
Follow GitHub repo instructions to install plugin
