Pushwanth

Led the retrieval and agent layer of an operational intelligence platform for production engineers, field operations, and maintenance planners. Built and deployed three operational agents (equipment troubleshooting with cited procedures, procedure search with revision control, and production issue analysis combining operational and incident context). Indexed 40,000 technical documents using Amazon Bedrock Knowledge Bases and Amazon OpenSearch with metadata filtering, reranking, and citation grounding; improved retrieval outcomes and reduced technical document lookup from 25 minutes to 4 minutes. Orchestrated agent workflows with LangGraph and Bedrock Agents, using MCP for controlled tool/data access and Agent2Agent for communication between independently deployed agents, with human approval for engineering sign-off paths. Migrated models from Claude 3.5 Sonnet/Haiku to Claude Sonnet 4.5/Haiku 4.5 by revalidating prompts, tool schemas, retrieval behavior, and citations against a golden

Engineered a fraud and AML decision platform scoring debit, ACH, and wire activity for investigators and analysts. Processed 3.2 million monthly transactions via Kafka ingestion and S3 storage. Built Airflow pipelines for data validation, feature preparation, training/scoring dataset creation, and batch orchestration. Developed behavioral and graph features (account/device/merchant/beneficiary) and implemented graph embeddings using Neo4j + GraphSAGE. Built classification models (Logistic Regression, Random Forest, XGBoost, LinearSVC and ensembles), optimized with class imbalance handling and hyperparameter tuning, improving PR AUC from 0.84 to 0.91. Tuned operating thresholds to reduce alert volume by 20% while maintaining baseline recall and reduced average investigation time from 40 minutes to 30 minutes. Deployed scoring via SageMaker endpoints and Kafka consumers; packaged APIs with FastAPI and Docker, published reason codes with SHAP explanations for explainability, and monitored

Developed the ML layer for prior authorization review and clinical document intelligence. Built data pipelines for X12 278 request/response transactions, FHIR resources aligned to Da Vinci prior authorization guides, and CCD documents using Airflow, Python/SQL, S3, and PostgreSQL. Implemented CCD and HL7 v2 ingestion to extract text/section structure, generate searchable metadata, chunk documents, and index in Amazon OpenSearch. Fine-tuned ClinicalBERT for authorization request classification and missing information detection; built BERT-based NER pipelines (BERT/spaCy/NLTK and healthcare dictionaries) and connected extracted entities to policy retrieval. Improved retrieval with metadata filters, document chunking, FAISS/OpenSearch vector search, domain embeddings, and cross-encoder reranking, reducing policy lookup time from 15 minutes to 3 minutes. Created and maintained an evaluation dataset and tracked runs and metrics using MLflow. Deployed classification and retrieval services vi