shital

Data Scientist

Fine-tuned GPT-3.5, GPT-4, and LLaMA models on proprietary datasets using Azure OpenAI and secure training environments; improved relevance and reduced hallucination through prompt optimization and evaluation.. Built OCR and computer-vision pipelines using Azure Computer Vision, PyTorch/ONNX, YOLO, Faster R-CNN, and U-Net to extract structured information from complex business documents.. Developed production MLOps workflows with MLflow, Azure Databricks, Delta Live Tables, Airflow, and Azure DevOps for training, evaluation, model versioning, retraining, and drift detection.. Developed predictive models using XGBoost and Random Forest with SHAP and LIME explainability for churn, claim-denial, utilization, and risk-oriented prediction use cases.. Applied LDA and BERTopic to large-scale complaint, appeal, and transcript data to identify emerging themes and actionable issues.. Built secure document ingestion and preprocessing pipelines using Unstructured.io for PDF, HTML, Markdown, and Office content used in RAG and knowledge-retrieval systems.. Applied LoRA and PEFT for parameter-efficient LLM fine-tuning in secured Azure ML and AWS SageMaker environments, reducing compute requirements for experimentation.. Implemented responsible AI controls including PII redaction, safety filters, audit trails, and output-quality evaluation for enterprise GenAI applications.. Delivered RAG-based summarization, document Q&A, and context-aware drafting solutions that increased documentation efficiency by more than 40%.. Conducted A/B testing across Azure OpenAI, AWS Bedrock, and Cohere to evaluate model quality, latency, cost, and privacy trade-offs.. Mentored three junior engineers on GenAI development, RAG architecture, MLOps practices, and responsible AI delivery.