Andrew Xu

I’m a Senior SWE (L5) at Google focused on building production agentic AI systems at scale. Over 10+ years, I’ve designed end-to-end agent pipelines that can plan, retrieve, act with tools, and learn in closed loops—reliably serving at massive interaction volumes.

At Google (Dialogflow CX), I architected a production multi-agent system with LangGraph, deploying a Supervisor Agent that dynamically routed work across specialized agents for intent resolution, knowledge retrieval, and compliance checking. This reduced dialog resolution latency by 38% and cut human escalations by 35% across billions of yearly interactions.

I also built hybrid RAG infrastructure using dense-sparse retrieval, BM25 re-ranking, and Vertex AI Vector Search over millions of enterprise documents—improving retrieval accuracy by 24% and enabling real-time knowledge base updates without full index rebuilds. For personalization, I designed a tiered long-term memory architecture spanning working memory, episodic storage, and semantic indexing, improving enterprise customer CSAT by 22%.

My approach emphasizes evaluation, safety, and observability: I engineered automated nightly evals with RAGAS and LangSmith to catch regression failures early, and I implemented Guardrails AI prompt-injection defense with structured output validation while maintaining sub-200ms p99 response latency. Previously at Motorola Solutions (Avigilon), I delivered closed-loop self-learning agents for video analytics, reducing model staleness from days to under 4 hours, and built autonomous Azure IoT fleet management that reduced mean time to remediation by 55%.