Background: Leading life insurance company wanted to digitize its claims triage process, which was heavily manual and prone to delays and human error. Fraud detection was reactive, and historical patterns weren’t being leveraged. They envisioned an AI-powered claims scoring engine to flag anomalies early and streamline investigations.

Solution Implementation:

1. Data Collection & Feature Engineering: We ingested 1.2 million historical claims records. Data was cleaned, normalized, and enriched with third-party sources (e.g., hospital ratings, policyholder history).

2. Model Development: We tested multiple algorithms (XGBoost, Random Forest, Logistic Regression) and selected XGBoost for its superior precision-recall tradeoff. Feature importance was reviewed with domain experts.

3. System Integration: Built using Python (FastAPI), the engine was integrated with clients’s claims platform via Kafka streams to enable real-time scoring. Each claim was scored in under 1.2 seconds.

4. OCR & NLP Add-on: Claims often involved scanned medical PDFs. We added a Tesseract-based OCR layer and custom NLP models to extract and interpret diagnosis and treatment patterns.

5. Visualization & Risk Dashboard: Developed an interactive dashboard in React with risk categories, investigator assignments, and timeline visualizations for each claim.

Results:

• Flagged 89% of risky claims with 92% accuracy.

• Reduced triage time by 70%.

• Enabled early detection of ₹3.8 Cr in fraudulent claims in 6 months.

• Scalable architecture deployed in 16 weeks, with future-readiness for underwriting expansion.