Kevin Phan

Senior AI Engineer with 8+ years of experience building production LLM applications, agentic AI systems, and RAG pipelines for healthcare platforms. Specialized in multi-agent orchestration, LLM evaluation, human-in-the-loop workflows, and reliable, cost-effective, compliant AI systems using AWS Bedrock and Azure OpenAI. Full-stack foundation in Python, Go, Node.js, React, and TypeScript, with a track record of taking AI products from prototype to production in regulated environments.

Senior AI Engineer at Abridge AI (2023-04 – 2026-06)

• Orchestrated multi-agent clinical workflows with LangGraph and MCP for agent-to-agent tool calling across scheduling, EHR updates, and prior authorization, using LangGraph interrupts to route high-risk actions to clinician approval and cutting manual processing effort by approximately 35%.
• Built LLM evaluation pipelines in LangSmith to track hallucination rate, retrieval precision, and response correctness across prompt and RAG changes, catching regressions pre-release and reducing production workflow failures ~30%.
• Hardened agent reliability with exponential-backoff retries (Tenacity), LangGraph checkpointing, and idempotent execution, enabling automatic recovery from transient LLM, STT/TTS, and EHR-API failures.
• Optimized LLM cost and performance using AWS Bedrock prompt caching and model routing, directing routine tasks to Claude Haiku and complex clinical reasoning to Claude Sonnet, reducing per-workflow token spend by approximately 25% without sacrificing response quality.
• Enforced compliance with AWS Bedrock Guardrails, schema-based output validation, and audit logging to keep generated clinical content within PHI-safe, HIPAA/FHIR-compliant bounds.
• Built a real-time voice AI pipeline with LiveKit, Deepgram (STT), and ElevenLabs (TTS), streaming transcription and synthesized responses over WebSockets into downstream clinical automation.
• Built RAG pipelines combining vector retrieval (pgvector) and keyword search to ground clinical notes, visit summaries, and discharge instructions in patient records.
• Built an internal AI observability dashboard (React + Chart.js) visualizing evaluation metrics, workflow success rate, and latency trends for the team.
• Mentored engineers through code review and architecture discussions, and used Claude Code and Cursor to speed up implementation while maintaining review rigor for a regulated environment.

AI Engineer at Ambience Health (2022-04 – 2024-03)

• Built RAG-based clinical summarization using Azure OpenAI and Azure AI Search over clinical documents and visit records, reducing documentation review time ~50%.
• Built a context-aware clinical chatbot on Azure OpenAI with FHIR-based retrieval, letting clinicians query patient records in natural language.
• Built a voice AI receptionist using Vapi and Azure Speech (STT/TTS) to automate patient intake and front-desk call handling.
• Built conversational workflows in Microsoft Copilot Studio, connecting front-desk and administrative processes to backend systems.
• Improved output reliability through structured output validation (JSON schema) and prompt refinement, reducing malformed responses ~30%.
• Implemented retrieval grounding to reduce hallucinations, improving factual accuracy ~20% in internal evaluation.
• Built backend APIs with Node.js/NestJS integrating Azure OpenAI and FHIR data sources for the clinician-facing React app.

Software Engineer at Zing Health (2021-06 – 2022-03)

• Built the enterprise Member Portal (React, Node.js, Go, PostgreSQL), serving 20K+ daily users with self-service access to benefits, claims, EOBs, and provider search.
• Built backend services for Care Coordination & Case Management, giving nurses and care teams tools to manage care plans and member tasks.
• Built Python data-processing pipelines supporting population health analytics and high-risk member identification.
• Deployed services on AWS (ECS, RDS, S3) and implemented CloudWatch monitoring and alerting to support scalable, HIPAA-compliant workloads.
• Optimized PostgreSQL queries and backend services, improving response times as member enrollment and traffic grew.

Software Developer at Google (2016-08 – 2021-04)

• Joined a cross-functional engineering team from early ideation through production launch of a new internal developer tool, contributing to problem framing and solution design alongside senior engineers.
• Built and shipped features in JavaScript, HTML/CSS, and Python that turned early prototypes into a production tool adopted across internal engineering teams.
• Built Python services and REST APIs automating internal engineering workflows, reducing repetitive manual work for engineering teams.
• Developed a strong foundation in production engineering rigor and ownership, focusing on user impact and product outcomes beyond implementation details, a mindset carried into subsequent AI engineering roles.

Bachelor's Degree in Computer Science – Seattle University (2012-08 – 2016-05)