HariPriya M

Built and productionized healthcare AI capabilities spanning clinical documentation extraction, ICD-10/CPT mapping with human review for uncertain cases, readmission risk modeling, and payer workflow automation using FHIR R4 and EDI X12. Engineered claims and remittance parsing (837/835), appeal document classification with evidence checks, and HIPAA-aligned PHI pipelines with encryption, minimum-necessary access, audit logging, tokenized identifiers, and controlled retention. Implemented FastAPI/gRPC microservices and streaming pipelines (Kafka/Spark) with replayable ingestion, quarantine queues, and monitored reliability. Developed PHI-safe observability and retrieval improvements using chunking, metadata filtering, semantic search, and reviewer-labeled relevance sets, plus LLM evaluation and CI/CD gates (MLflow, Docker, GitHub Actions) for governed releases into clinical/risk workflows.

Designed and built governed lending and risk assistant workflows using LangGraph/LangChain and AutoGen, including policy retrieval, tool-calling underwriting services, stateful exception handling, and human approvals. Deployed production LLM services on Vertex AI endpoints and Cloud Run with secure workload identity, regional controls, and autoscaling. Implemented enterprise RAG with Vertex AI Vector Search, BM25 hybrid retrieval, metadata filtering, and reranking; added structured fraud narrative generation with schema validation and prohibited-content checks. Built evaluation suites that measure retrieval recall/groundedness/tool completion/latency/cost and block regressions during CI. Engineered typed, async FastAPI/gRPC services connected via Pub/Sub and Cloud Functions, with caching, observability (LangSmith/MLflow/Datadog), and Terraform-based secure networking (VPC/IAM).

Delivered e-commerce machine learning solutions including customer recommendation and ranking (collaborative filtering in PyTorch), demand forecasting (XGBoost/LSTM/ARIMA baselines), search relevance/ranking using scikit-learn and spaCy, churn modeling with calibration and explainable drivers, and fraud detection using XGBoost/Isolation Forest on PySpark. Built event-driven pipelines with Kafka/Spark Streaming for clickstream feature generation, supply chain risk models, and support-ticket intent classification pilots with BERT. Established MLflow tracking, scheduled model refreshes with Airflow, containerized inference services for deployment (SageMaker), and implemented caching with Redis to support seasonal traffic. Developed robust data validation and automated testing using PyTest and CI workflows with Jenkins/Git.