Sai Prasanth Kaligotla

Predicted fraudulent insurance claims by engineering graph-based features (network centrality, temporal claim velocity) and training Graph Neural Networks (PyTorch Geometric), improving fraud detection precision by 14% and preventing an estimated $3.2M in projected quarterly losses. Built data quality validation and outlier detection pipelines on terabytes of transactional claims data using advanced SQL window functions and complex joins. Developed a transformer-based NLP classification pipeline for document categorization, addressing class imbalance with SMOTE and achieving 92% accuracy while reducing manual review effort by 35%. Led end-to-end modeling for 5M+ active policies from feature selection to hyperparameter optimization (reducing training time by 42%) and implemented data drift monitoring with SageMaker Model Monitor to sustain ROC-AUC performance. Partnered with risk stakeholders to set actionable decision thresholds using precision-recall tuning.

Built fraud detection models by engineering graph-based features (e.g., network centrality, temporal claim velocity) and training Graph Neural Networks with PyTorch Geometric, improving fraud detection precision by 14% and preventing $3.2M in projected quarterly losses. Implemented transformer-based NLP classification pipelines for underwriter document triage, handling class imbalance with SMOTE and achieving 92% accuracy while reducing manual review by 35%. Led an end-to-end modeling lifecycle for 5M+ active policies including feature selection, hyperparameter optimization, and data drift monitoring with SageMaker Model Monitor to sustain ROC-AUC. Translated model probabilities into business rules by tuning decision thresholds using precision-recall tradeoffs, and developed a RAG framework to reduce policy retrieval latency to sub-250ms.

Delivered enterprise churn models using Kaplan–Meier survival analysis on 10M+ cross-channel user records to support retention campaigns and preserve $2.1M in at-risk contracts. Built Python preprocessing pipelines to handle missing data and skewed distributions, validating statistical significance with chi-square tests before training. Engineered TF-IDF and behavioral features from 50K+ weekly unstructured tickets to improve routing with a baseline Logistic Regression and an optimized XGBoost model, reaching 92% routing precision and reducing annual operational costs by $120K. Reduced unplanned manufacturing downtime by 22% using rolling-average feature engineering and correlation analysis on 500GB IoT sensor data, forecasting failure risks 48 hours ahead with predictive maintenance modeling. Conducted ablation studies and SHAP-based explainability, then used A/B testing frameworks to confirm performance before production deployment.