Aruna Fnu

Led agentic AI and GenAI-assisted workflows to modernize fraud detection, dispute resolution, and regulatory responses for a global payments network. Architected autonomous multi-agent systems and retrieval-augmented generation pipelines, plus enterprise document-processing capabilities to accelerate investigator decision-making. Built RAG chatbot and conversational assistant services with streaming, retrieval improvements (hybrid BM25 + dense embeddings, cross-encoder reranking), and GraphRAG over a Neo4j knowledge graph. Implemented human-in-the-loop checkpoints, prompt/version governance, evaluation frameworks (RAGAS, DeepEval, TruLens, LLM-as-judge), and production guardrails including PII masking and prompt-injection testing. Delivered APIs and containerized services on Kubernetes with observability via OpenTelemetry, LangSmith/Langfuse, and Azure monitoring.

Designed and deployed large-scale ML systems for personalization and demand forecasting across retail and e-commerce. Built recommendation and ranking models (collaborative filtering, XGBoost/LightGBM) serving 100M+ monthly active customers and improving click-through rates. Developed real-time feature pipelines with Kafka, PySpark, and a Feast-based feature store, and deployed low-latency inference services with FastAPI/BentoML/KServe on Kubernetes. Delivered forecasting improvements using Prophet/ARIMA and gradient-boosted models, reduced forecast error (MAPE) by 22%, and implemented experimentation and A/B testing with MLflow and statistical significance testing. Added semantic search using embeddings + FAISS, and implemented monitoring/drift detection with Evidently/Prometheus/Grafana/Datadog.

Built NLP, transformer-based text classification, semantic search, and predictive modeling solutions for claims and clinical documentation. Developed document-processing pipelines (OCR/parsing) using Azure Document Intelligence, PDF tools, and PHI-aligned handling with Presidio-based redaction. Implemented semantic retrieval over historical claims using embeddings with FAISS and Elasticsearch, and prototyped Q&A capabilities to surface relevant evidence and policy language, reducing manual review time. Fine-tuned BERT-style models for clinical tasks, combined structured EHR fields with NLP-extracted entities, and created high-risk claim scoring models with explainability (SHAP). Deployed models using Vertex AI Pipelines/Endpoints and served via FastAPI; mentored engineering best practices and CI/CD for reliable releases.