Vinay Dontri

Senior Data Scientist and AI Engineer specializing in production grade Generative AI and Agentic systems, with 10+ years spanning enterprise software engineering, machine learning, and applied AI across banking, government, and healthcare. Built and shipped Retrieval Augmented Generation (RAG) systems, multi agent architectures (planner executor workflows, Model Context Protocol, Agent Development Kit, Agent to Agent communication), and multimodal pipelines (Whisper, CLIP, OCR) for real world use cases: symptom triage, demographic forecasting, fraud detection, and enterprise knowledge search. Deep experience across the ML lifecycle, from Graph Neural Network fraud detection and time series forecasting to LLMOps (RAGAS, LangFuse, MLflow) and secure, compliant deployment (HIPAA, SOX, SOC2) on both AWS and Azure.

Started in enterprise Java engineering and cloud migration before moving into machine learning and AI, bringing a strong production systems foundation to the data science work. Partners directly with compliance, clinical, and operational stakeholders to take AI systems from prototype to production in regulated environments.

Senior Data Scientist (Contract) at Centene (2026-03 – Present)

Project: RAG Powered Symptom Triage and Care Navigation. A healthcare AI platform that helps patients describe symptoms in plain language and routes them to the right level of care.

• Built a Retrieval Augmented Generation system grounded in clinical triage protocols using Azure OpenAI's GPT 4o for structured clinical reasoning, with Anthropic's Claude as a secondary safety reviewer on sensitive recommendations, improving triage recommendation accuracy by 28% over the prior rules based logic.
• Designed a LangGraph/LangChain conversational agent handling 15,000+ patient symptom intake conversations a month with an average end to end response time under 6 seconds.
• Fine tuned a CLIP vision encoder in PyTorch to triage visual symptoms (rashes, wounds) from patient submitted photos, correctly flagging over 90% of cases needing in person evaluation.
• Added voice based intake (Whisper) and OCR (AWS Textract) to auto extract data from intake forms and insurance cards, and built the supporting data pipelines (Python/SQL), OpenSearch retrieval index, and FastAPI microservices behind AWS API Gateway.
• Set up confidence thresholds that route low confidence cases to a human nurse line, containerized the platform with Docker, and maintained 99.9% uptime via CloudWatch, all within HIPAA requirements (IAM, RBAC, KMS encryption, FHIR/HL7 alignment).
• Partnered with clinical stakeholders, care navigators, product, and engineering teams to validate triage recommendations against real clinical judgment and iterate the flow based on patient outcomes.

Senior Data Scientist (Contract) at State of Michigan (2024-09 – 2026-02)

Project: Statewide 2050 Demographics and Population Projections. Forecasting Michigan's population to 2050 at the county and regional level.

• Built Prophet based time series forecasting models, benchmarked against Scikit learn/XGBoost regressions and a TensorFlow/Keras LSTM, achieving under 4% mean absolute percentage error on 5 year backtests before generating county level projections through 2050 for all 83 Michigan counties.
• Designed a RAG conversational agent (initially on Meta's Llama 3, later rebuilt around MCP and Google's Agent Development Kit) that cut the time for planners to get a grounded answer from days to seconds.
• Implemented Agent to Agent communication across retrieval, forecasting, and reporting agents, and used RAGAS/LangFuse to monitor answer groundedness, cutting manual QA review time by about 40%.
• Unified Census Bureau, American Community Survey, and state vital records data into a single Python/SQL pipeline processing over 20 million historical records, and used AWS Textract and Whisper to digitize historical documents and transcribe public hearings.
• Applied CLIP based image classification to land use imagery to model urban growth patterns, containerized the forecasting and agent services with Docker on AWS, and kept pipeline uptime above 99.5% via CloudWatch.
• Built Power BI and Streamlit dashboards presenting findings to state planning officials, supporting infrastructure, housing, and resource allocation decisions affecting roughly 10 million residents.

Data Scientist (Contract) at Truist Bank (2020-08 – 2024-06)

Project: Synthetic Identity Fraud Detection using Graph Neural Networks. Modeling customers, accounts, devices, and transactions as a connected network to catch fraud rings that look legitimate individually.

• Designed and trained GraphSAGE and GCN models in PyTorch Geometric on a Neo4j entity graph tied to an estimated $8M+ in annual attempted fraud losses, improving detection recall by roughly 32% over the prior rule based system.
• Used SHAP to explain a downstream gradient boosted classifier combining GNN embeddings with engineered features to fraud and compliance teams in a SOX/SOC2 governed environment, and tracked experiments in MLflow, reducing false positive rates by about 20%.
• Processed over 500 million transaction and account records with PySpark on AWS EMR, orchestrated with Apache Airflow and staged in S3, then trained the PyTorch Geometric models as bring your own container jobs on AWS SageMaker.
• Deployed the fraud models as scoring services on Docker/Kubernetes (EKS), provisioned with Terraform and shipped through GitHub Actions CI/CD, cutting new account fraud review time from hours to near real time and putting the $8M+ in previously missed fraud exposure under active detection.
• Built Tableau dashboards tracking fraud trends, model performance, and case outcomes for leadership and compliance across a multi billion dollar transaction portfolio.
• Partnered closely with fraud analysts, compliance, and engineering teams to get the models into production and fit them into the bank's existing investigation workflow.

NLP Engineer at Microsoft (2018-01 – 2020-07)

Project: OmniContext, Enterprise Conversational Search and Document Intelligence (Digital Transformation / Enterprise Search division). Turning unstructured, multilingual enterprise data into a searchable knowledge base and Q&A agent.

• Built NLP pipelines in Python using spaCy and custom entity extraction rules, structuring over 2 million unstructured emails, chats, and PDFs per month across 5+ languages.
• Layered a custom BERT embedding similarity index on top of Azure Cognitive Search for semantic reranking, improving top 3 search relevance by about 25% over the keyword only baseline.
• Trained intent and entity recognition models in Microsoft LUIS and built QnA Maker knowledge bases connected to the Azure Bot Framework, powering a Q&A bot that resolved over 60% of employee queries without human escalation.
• Built multi language support with the Microsoft Translator API and OCR pipelines using Azure Computer Vision to digitize scanned PDFs, cutting manual document processing time by about 40%.
• Automated data ingestion and indexing with Azure Data Factory and Azure Functions, storing structured outputs in Cosmos DB for sub second query response.
• Containerized the services on Azure Kubernetes Service with CI/CD via Azure DevOps, supporting over 10,000 daily active users, and tracked performance with Application Insights.