Patrick Yuen

Built secure enterprise GenAI and agentic-AI capabilities for banking operations, risk, and technology teams. Owned RAG design, agent orchestration, evaluation, production reliability, and cloud deployment across high-governance workflows. Designed multi-agent investigation workflows that decompose requests, gather evidence, compare findings, and produce cited recommendations for human review. Engineered hybrid retrieval using embeddings and keyword search with metadata filters, query rewriting, evidence-aware reranking, and context deduplication. Implemented governed LLM tool calling for case search, policy retrieval, workflow status, service telemetry, and approved actions with scoped permissions and audit records. Established evaluation for retrieval recall, groundedness, accuracy, hallucinations, task completion, latency, and cost. Improved grounded-answer accuracy by 24% and reduced end-to-end latency by 31% via chunking, reranking, citation validation, parallel tool execution, ca

Built production insurance ML and document-intelligence services that transformed forms, correspondence, claim events, and policy data into structured features, classifications, prioritization signals, and recommendation workflows. Built Kafka and Spark pipelines to ingest claim events, scanned documents, correspondence, and policy metadata for model-ready features. Developed document extraction and classification models for form type and coverage attributes, claim intent, missing information, and processing priority. Trained NLP classifiers for topic, urgency, sentiment, and routing of adjuster notes and customer correspondence. Created recommendation and ranking services to prioritize processing queues and surface related policies and prior cases. Improved routing precision by 19% using layout-aware features, hard-negative sampling, calibrated thresholds, and error analysis with claims specialists. Standardized experimentation and reproducibility with MLflow (model registry, approval

Modernized Prudential policy-servicing and claims integration across legacy and cloud systems, owning APIs, event-driven services, AWS migration, resilience patterns, performance optimization, and operational tooling. Owned Java and Spring Boot APIs for policy lookup, customer updates, document status, claim intake, and servicing workflow actions. Designed event-driven integration services using Amazon Kinesis, SQS, and SNS to process policy and claim events across operational systems. Integrated rules and risk-scoring outputs into low-latency, versioned APIs with validation, explainability metadata, and fallback logic. Reduced p95 API latency by 37% using Redis caching, DynamoDB access-pattern redesign, request coalescing, and parallel downstream calls. Migrated services to AWS with ECS, Lambda, API Gateway, RDS, DynamoDB, S3, and CloudWatch, managed with Terraform. Improved resilience with idempotent consumers, circuit breakers, bounded retries, dead-letter recovery, health checks, a