AI-powered insurance claim intake automation using Azure Document Intelligence and GPT-5 for extraction, policy matching, and confidence-scored coverage analysis.

Risk Profiler
Machine learning prototype that profiles fraud risk in insurance claims with explainable, auditable predictions
Highlights
Insurers face heavy financial exposure from fraudulent claims. Manually reviewing every claim is costly and impractical, while blanket approvals risk paying out fraud.
Build an ML prototype that scores fraud risk per claim with a configurable threshold, explains each prediction transparently, and supports investigation workflows.
Trained an AutoGluon ensemble on the Travelers NESS Statathon dataset, added SHAP per-feature contributions and GPT-4o-mini summaries, surfaced via an interactive dashboard.
Instant fraud probability with auditable, additive reasoning, letting analysts focus on high-risk and edge cases instead of scoring every claim by hand.
Core Team
Overview
Risk Profiler is a machine learning prototype that helps claims analysts identify potentially fraudulent insurance claims, prioritize investigations, and reduce financial losses. Rather than reviewing every claim by hand, analysts get an instant fraud probability for each case along with clear reasoning that supports their decisions and regulatory compliance.
The system pairs a trained AutoGluon ensemble with SHAP explainability and optional LLM-generated summaries. Every prediction is fully transparent: SHAP contributions show exactly how each feature pushed the risk score up or down, following a simple additive form: baseline plus feature impacts equals the final risk score. An interactive dashboard lets analysts adjust claim signals and explore "what-if" scenarios in real time.
Data
The model is trained on the 2023 Travelers NESS Statathon dataset of synthetic insurance claim records with fraud labels, covering driver demographics, claim details, and vehicle information:
- •Travelers NESS Statathon: Open 2023 Kaggle competition data with labeled fraudulent and legitimate claims, making the prototype reproducible and auditable.
- •Claim Features: The 10 most impactful signals by SHAP: annual income, age of driver, claim day of week, higher education, past number of claims, safety rating, witness present, gender, estimated payout, and living status.
Claims exceeding a configurable 65% fraud-probability threshold are flagged as high risk, focusing analyst attention where it matters most.
Methods
Risk Profiler runs a prediction pipeline that combines a trained AutoGluon model with SHAP explainability and optional LLM summaries:
- Stage 1 (Prediction): An AutoGluon TabularPredictor (NeuralNetTorch ensemble) outputs a fraud probability and a binary decision against the configurable threshold.
- Stage 2 (Explainability):
- •SHAP Contributions: A KernelExplainer with a 25-sample background computes per-instance feature contributions.
- •Additive Scoring: Baseline plus the sum of feature impacts reconstructs the risk score, so each input's push up or down is explicit.
- Stage 3 (Summary & UI): An OpenAI model (GPT-4o-mini, configurable) generates a natural language assessment, with a template fallback when no API key is set, and the dashboard renders the risk score, summary, and an explainability panel with the additive contributions.
Results
By replacing manual scoring with explainable, on-demand predictions, the prototype delivers practical gains for fraud investigation workflows:
- •Review Speed: Instant fraud probability and explanations replace full manual scoring and analysis.
- •Analyst Focus: Reviewers concentrate on high-risk and edge cases instead of examining every claim.
- •Consistency: Standardized scoring and reasoning remove reviewer-to-reviewer variance.
- •Explainability: SHAP contributions and an LLM summary replace hand-written notes with auditable reasoning.
A walkthrough of scoring an auto claim, adjusting claim signals to see the fraud probability cross the 65% threshold, and inspecting the SHAP drivers and summary is shown in Figure 1, and the full dashboard layout is shown in Figure 2.
Conclusion
Risk Profiler shows how explainable machine learning can support, rather than replace, human fraud analysts. By pairing accurate AutoGluon predictions with transparent SHAP contributions and concise LLM summaries, it turns opaque risk scores into auditable, defensible decisions while keeping a human firmly in the loop.
As a prototype, it is built for extension: the decision threshold and the OpenAI summary model are both configurable, and the batch-ready API is positioned for integration into existing claims platforms and larger fraud-detection workflows.
Keep Exploring
Non-invasive ICP monitoring using deep learning on cerebral blood flow signals from near-infrared photonic sensors.



