vishal.singh

◦ Developed and deployed forecasting modules using anomaly detection and GenAI for the Cost of Care platform — converted large-scale claims data (60M+ members) into 3-month, state-level predictions to support proactive interventions. ◦ Designed Bayesian hierarchical models to quantify uncertainty in cost-of-care forecasts, producing calibrated credible-interval outputs that improved pricing decisions and cut forecast error variance by 12%. ◦ Worked on NLP-based feature engineering and optimisation using financial, clinical, and demographic data to help reduce underwriting turnaround by 20% and improved real-time rating accuracy, contributing to projected $1.5M in projected cost savings by 2025. ◦ Built probabilistic models to flag high-cost claimants and special medical conditions — enabled $300K in annual savings via improved cost visibility and risk stratification. Increased condition identification coverage from 48% to 81%, supporting a 15x expansion in outreach. ◦ Created an NL-to-SQL pipeline using fine-tuned LLMs, semantic search, and visual query flows — achieved 95% accuracy and improved performance by 31% over existing methods. ◦ Co-developed a clinical QA assistant using LLAMA2, LangChain, and a QLoRA-tuned RAG pipeline — enabled secure, domain-specific medical question answering at scale. ◦ Automated manual ticket triage workflows using statistical classifiers (Naive Bayes, XGBoost) with Word2Vec+BERT embeddings — used probabilistic baselines to guide model iteration and achieved effort savings of 8–10 FTEs across support teams. ◦ Contributed to a 40% improvement in underwriting throughput by building a clinical rules engine, integrating ML-based risk scoring, and enabling LLM-assisted workflow automation — delivered $100K+ in operational savings. ◦ Supported deployment of ML models across AWS, Azure, and on-prem environments using Docker pipelines, ensuring stable rollout of real-time APIs and dashboards.

◦ Built a RAG-style semantic search QA chatbot using BERT embeddings and vector search over 100K+ files — enabled fast, context-relevant answers via semantic similarity, without using LLMs. ◦ Mentored 3 interns and 3 junior developers on ML workflows, code quality, and model deployment — improved maintainability and established reusable standards across the team.

◦ Industry Collaboration (GreatFour Systems): Conducted independent research on point cloud segmentation, graph neural networks (GNNs), and few-shot learning under faculty guidance — developed core model architectures for structured reasoning and sparse data classification in industrial settings. ◦ Academic Research: Designed and evaluated experimental GNN-based models for knowledge graph completion, link prediction, and representation learning under low-resource constraints.