Bhavani Ganipisetti

Designed the core workflow for an internal Generative AI knowledge assistant that helps banking users search approved policies, procedures, and product documentation via conversation. Built an enterprise RAG architecture with separate ingestion, chunking, metadata, embeddings, retrieval, prompt construction, response generation, and validation components using Azure OpenAI and Azure AI Search. Implemented metadata filtering by document type, business area, and access context to narrow retrieval and prevent exposure of sensitive enterprise information, and tuned vector/semantic/hybrid search to improve ranking quality. Used LangChain and LangGraph for conditional routing, validation nodes, approved tool invocation, and human-in-the-loop review workflows. Exposed GenAI capabilities to applications through controlled FastAPI REST services, adding security controls (authentication/authorization/entitlements) aligned with risk and application teams rather than relying only on prompt instruc

Built document-processing and AI workflows for insurance claims and policy operations involving PDFs, scanned documents, forms, correspondence, and structured application data. Split solutions into ingestion, OCR/extraction, validation, preprocessing, classification, information extraction, business validation, and API integration stages. Used Azure Form Recognizer / Azure AI Document Intelligence for OCR/extraction and added Python validation to account for OCR inaccuracies. Developed preprocessing logic to normalize extracted content and reduce variability from inconsistent document formats and missing fields. Created NLP pipelines for document classification and information extraction, choosing Scikit-learn and XGBoost for explainable, predictable classification and prioritization. Built reusable feature-engineering and validation logic in Python for production scoring, and used PySpark/Azure Databricks for large-scale transformations. Tracked experiments and managed reproducible mo

Converted citizen-service demand and inquiry questions into classification, forecasting, and NLP/analytics problems in support of planning and operations. Designed reusable data-preparation workflows in Python/SQL using Pandas and NumPy for profiling, cleaning inconsistencies, feature creation, and building training datasets. Developed classification models with Scikit-learn and XGBoost for service-request categorization, comparing approaches using validation performance and usability. Built forecasting workflows using historical trends and statistical methods to support planning teams with forward-looking workload estimates. Implemented NLP workflows with BERT and Hugging Face Transformers for free-text citizen inquiries not reliably captured in structured fields, evaluating models with precision, recall, F1-score, ROC-AUC, MAE, and RMSE and communicating tradeoffs. Used Spark/PySpark when distributed processing was needed and created scheduled batch-scoring plus Flask/FastAPI service