Work Experience

Smart Triage Agent Runtime: Designed an end-to-end ReAct-based AI agent runtime for autonomous driving root-cause analysis using LangGraph and LangChain, achieving 81.6% benchmark accuracy on 1000 real-world incidents.
Cost-Efficient Context Engineering: Reduced LLM inference cost by 97% ($3–4 to $0.09 per incident) through prompt caching, context compression, and dynamic skill retrieval.
Agent Skill Framework: Architected an Agent Skill / Macro Function framework that transforms domain knowledge into executable tools with structured metadata, enabling grounded reasoning through dynamic vector retrieval.
Production Triage Platform: Independently architected and productionized an end-to-end AI triage platform from agent runtime to cloud deployment (interactive IDE and offline message-queue workers on Kubernetes); led cross-functional integration with Perception and Planning teams, and mentored 1 intern and 1 new graduate engineer.
No-Comms Pipeline: Designed a lightweight intra-process messaging mechanism to bypass shared queues, improving memory efficiency by 65.45% and enhancing system responsiveness.

Built a cross-team visualization platform that unified simulation and real-world driving data, improving perception debugging efficiency by 40%.
Developed OpenCV-based LiDAR visualization tools, enabling engineers to identify subtle sensor anomalies during perception model evaluation.

Engineered a scalable browser automation framework supporting 40+ e-commerce platforms, reducing regression testing effort by over 90%.