AI-Driven Fraud Case Summarizer (Risk Explainability)
An AI-assisted fraud analytics system that combines model risk outputs, SHAP-based explainability, and GenAI narratives to transform raw transaction data and fraud signals into structured, analyst-ready case summaries—supporting faster, more transparent fraud review without automating final decisions.
Tags & Technologies
Key Impact & KPIs
- Consolidates risk scores, explainability signals, and narratives into a single case view
- Reduces analyst effort spent interpreting model outputs and raw features
- Improves transparency of fraud risk drivers through SHAP-based explanations
- Standardizes fraud case documentation with structured summaries
- Demonstrates responsible human-in-the-loop GenAI for regulated environments
Project Overview
1. End-to-End Fraud Case Summarization Pipeline
Built an end-to-end fraud case summarization pipeline that ingests transaction-level data, engineered risk features, and model outputs (risk score and probability) and converts them into structured, human-readable fraud case summaries aligned with analyst workflows.
2. Model Explainability with SHAP
Integrated model explainability using SHAP by surfacing top contributing risk features, feature–impact pairs, and summarized SHAP reasoning—ensuring analysts can understand why a case was scored as risky, not just the final score.
3. Structured Input–Output Schema
Designed a structured input–output schema that separates raw inputs (input_text), predictive signals (risk_score, risk_score_prob), explainability artifacts (top_shap_features, shap_summary), and AI-generated narratives (generated_narrative), enabling clear traceability and audit-friendly case reviews.
4. GenAI Narrative Layer
Implemented a GenAI narrative layer that synthesizes SHAP explanations, transaction behavior, and customer context into concise, plain-English fraud summaries—bridging the gap between numerical model outputs and analyst interpretation without influencing core risk scoring.
5. Interactive Streamlit Demo
Delivered an interactive Streamlit demo showcasing the full before-and-after workflow—from raw case inputs and explainability signals to AI-generated fraud narratives and recommended actions—demonstrating how predictive models and GenAI can be combined responsibly in fraud operations.
Model Selection Rationale
- Models/LLMs used: XGBoost / tree models + SHAP for reason codes; Flan-T5-Small for constrained narrative generation.
- Explainability & auditability: SHAP gave analyst-friendly reason codes; tree models are traceable for investigations.
- Controlled language: Flan-T5-Small was chosen for reliable, template-friendly summaries that avoid creative hallucination.