Architected an enterprise document intelligence solution achieving 92% average extraction confidence and 40% reduction in manual data entry, by designing Doxtract-IDP using AWS Bedrock multi-modal models (Claude Sonnet 4) with a serverless Lambda-SQS-DynamoDB event-driven architecture.

Implemented production-grade document processing with 90%+ field-level extraction accuracy via human-in-the-loop validation and post-processing rules, deployed on serverless AWS infrastructure handling 60600 second processing windows per document.

Designed and deployed high-throughput microservices on AWS EKS with horizontal pod autoscaling, load balancing, and CI/CD pipelines, delivering 99.8% system availability while handling 300+ concurrent requests during peak usage.

Architected an event-driven GenAI pipeline using LangGraph agent systems with persistent memory and chain-of-thought reasoning, improving response relevance scores by 20% and reducing manual policy review time by 15%.

Engineered a custom vector database retriever with LLM-based reranking, increasing search relevance scores by 25% and achieving 87% precision across 2,000+ synthetic insurance risk scenarios.

Optimized GPU infrastructure costs by migrating from local Ollama deployments to AWS Bedrock API integration, eliminating on-premise GPU dependencies and reducing monthly EC2 spend by 60%.

Deployed a self-hosted OCR service using Vision Transformer (Phi3.5 Vision) on EKS with Ray clusters for distributed computing, reducing monthly infrastructure costs by 20%.

Built a RAG pipeline with Kafka-based document ingestion and Redis query caching, implementing custom retrieval and reranking components with structured claim analysis to achieve 87% precision on insurance risk assessment across 2,000+ test scenarios.