AI/ML Engineer, Healthcare AI at Molina Healthcare

Reduced prior-authorization clinical review time from 47 minutes to 8 minutes by deploying a LangChain + Azure OpenAI GPT-4 assistant with hybrid dense/BM25 retrieval and cross-encoder reranking. Helped prevent approximately $4.2M in annual fraudulent Medicaid claim payouts across five state programs by architecting an XGBoost/scikit-learn fraud-detection pipeline on Azure Machine Learning. Enabled proactive intervention for 120,000+ high-risk Medicaid and Medicare members using an agentic risk-stratification workflow with LangGraph, Apache Spark, and gradient-boosting models. Improved data freshness to under 24 hours for six production AI systems by consolidating HIPAA-compliant feature pipelines across 14 clinical data sources in Databricks Delta Lake. Reduced ML deployment lead time from 3 weeks to 2 days using Azure DevOps CI/CD with MLflow experiment tracking, model monitoring, and data-drift detection. Strengthened governance and auditability across three AI workflows via explain

AI/ML Engineer at BNY Mellon

Automated extraction from 85,000+ regulatory filings per month by developing a transformer-based document-intelligence pipeline with LayoutLM and Hugging Face Transformers. Reduced manual document-review effort by approximately 1,400 hours by combining document preprocessing, transformer-based extraction, and structured-output validation. Reduced inference latency from 2.3 seconds to 140 milliseconds per request by migrating legacy ML scoring models to AWS SageMaker and applying ONNX graph optimization and TensorRT acceleration. Enabled natural-language search for 240+ financial analysts by integrating sentence embeddings, FAISS/Pinecone vector search, semantic retrieval, and reranking into an enterprise policy-search platform. Reduced analyst report-digestion time by 68% by fine-tuning Llama 3 with QLoRA and deploying a standardized 10-K and 10-Q summarization workflow through AWS Bedrock. Supported p99 inference latency under 200 milliseconds for 50,000+ concurrent prediction request