Anomaly Detection & MLApr 2024 – Jul 2024

Hybrid Fraud Detection System

Multi-layered anomaly detection combining deterministic heuristic rules and LLM-based signals

Hybrid PipelineHeuristic RulesLLM SignalsAnomaly DetectionHuman-in-the-LoopFinancial Risk

System Architecture & Challenge

Financial fraud detection faces a delicate trade-off between strict rule-based triggers and machine learning classifiers. Strict deterministic rules produce excessive false positives that frustrate legitimate users, whereas pure black-box classifiers often lack transparent interpretability required by financial auditing bodies.

This project developed a hybrid fraud detection pipeline that unifies deterministic heuristic rules with LLM-based reasoning signals to flag anomalous banking transactions while preserving operational clarity.

Multi-Layered Architecture

Heuristic Validation:Deterministic checks for IP legitimacy (public, private, invalid), IP change velocity, account age thresholds, and fund sufficiency.

LLM-Based Signals:Contextual reasoning signals analyzing irregular transaction patterns and generating natural-language risk justifications.

Human-in-the-Loop:Borderline risk cases routed automatically to an administrative audit interface for human review before final execution.

Verified Key Scope

Built a multi-layered hybrid architecture integrating deterministic business rules with machine learning and LLM-assisted reasoning signals.
Engineered transactional feature representations including IP classification (public/private/invalid), account age binning, and IP change velocity ratios.
Implemented human-in-the-loop review protocols for borderline risk scores, ensuring high-risk decisions maintain interpretability and auditability.
Provided explainable risk explanations detailing exact anomaly factors for flagged transfers.