Billy Y

About

LLM/RAG practitioner building clinician-facing AI search and Q&A inside EHR workflows, focused on trust, latency, and safety (grounded answers with citations, PHI controls, encryption/audit logs). Demonstrated real-time incident response for production LLM systems (e.g., fixing a metadata-filter deployment regression to prevent irrelevant results/cross-patient leakage) and strong demo/enablement skills for mixed technical and clinical stakeholders; also shipped a multi-model RAG tool at OrbeX Labs with upload/search/audit features for day-to-day adoption.

Experience

Education

Key Strengths

• Takes LLM prototypes to production with strict grounding and approved-source retrieval
• Builds trust in clinical AI systems via citations, evidence-based answering, and safety fallbacks
• Performance tuning for LLM/search systems (index/query optimization, caching to meet latency targets)
• Strong incident-style debugging of LLM/agent workflows using reproducible prompts, traces, and log analysis
• Rapid real-time remediation with root-cause analysis (e.g., metadata filter regression causing wrong time-window retrieval)
• Implements privacy/security controls for clinical data (PHI protections, access control, encryption, audit logs)
• Effective technical communicator—tailors demos/workshops for mixed clinician + engineer audiences with clear integration steps
• Partners with sales/customer teams to de-risk adoption through POCs, integration documentation, and fast post-launch iteration
• Designed end-to-end OCR→embeddings→retrieval pipeline for document/image processing
• Improved pipeline robustness via normalization and stronger error handling for inconsistent document formats
• Reduced end-to-end processing time significantly by introducing async processing
• Deployed and scaled microservices on AWS EKS using Docker + Helm with health probes
• Implemented GitOps CI/CD with GitHub Actions + ArgoCD (PR-based changes, automated sync/rollback)
• Led phased migration from VM-based services to containerized workloads on AWS ECS with low downtime
• Built/optimized event-driven pipelines with Kafka/RabbitMQ; mitigated spikes via partitioning and consumer concurrency
• Implemented idempotent processing to prevent data inconsistency during retries