Naveen Kumar

Built a Generative AI care-management assistant on Amazon Bedrock with Claude, using LangChain RAG over Pinecone to ground responses in approved clinical documents. Implemented multi-agent LangGraph workflows with stateful routing and human-in-the-loop checkpoints for clinical case triage and prior authorization support. Fine-tuned Llama 3.1 with LoRA on de-identified clinical text, and exposed LLM/ML services through FastAPI with Pydantic schemas and OAuth2. Applied Bedrock Guardrails and PHI redaction using Amazon Comprehend Medical for HIPAA compliance. Instrumented and evaluated prompts, retrievals, and agent runs with Arize Phoenix and ran RAG evaluations using RAGAS before releases. Reduced inference spend with caching and prompt/route optimization, and supported document and entity extraction pipelines using spaCy and Transformers. Deployed model training and pipelines on SageMaker with MLflow registry, and delivered evaluation dashboards via Streamlit and Tableau.

Developed a conversational assistant on Anthropic Claude via Amazon Bedrock to help engineers query suspension test data in natural language, reducing manual weekly analysis. Built RAG workflows with Bedrock Knowledge Bases, LangChain, and FAISS using failure-pattern documents for higher answer accuracy. Created an agentic document retrieval assistant with LangChain tools and memory to streamline internal Q&A. Designed serverless AI pipelines with Lambda, Step Functions, and API Gateway using AWS CDK, and served inference with FastAPI microservices on ECS secured via IAM and API Gateway authorizers. Automated document processing with Amazon Textract and Claude-based field extraction, and built an evaluation framework in Python to compare accuracy, groundedness, latency, and cost across Bedrock models. Contributed to AWS architecture for active-suspension preview/control workloads and developed a computer-vision inspection model on SageMaker to automate quality validation. Built forecas

Built a fraud detection platform on Amazon SageMaker for high-volume daily card transactions, improving performance and reducing false alerts and analyst review load. Implemented real-time inference services using FastAPI with Pydantic validation, deployed with Docker on Kubernetes for high-throughput scoring. Engineered low-latency streaming feature pipelines with Kafka and Spark Structured Streaming to compute behavioral fraud signals. Trained gradient-boosted models (XGBoost/LightGBM) handling class imbalance with SMOTE and using cost-sensitive learning and threshold tuning. Delivered explainability with SHAP and reason-code generation for model governance and regulatory review. Built semantic search over compliance documents using sentence-transformer embeddings stored in PostgreSQL with pgvector, and created NLP pipelines with spaCy and Transformers for AML/KYC document review and adverse-media screening. Established MLOps with MLflow, GitHub Actions CI/CD, PyTest, and monitoring