Three-plus years at AWS — Amazon Lex and Lambda. Features owned end-to-end, design through multi-region launch. Built for reliability and operational efficiency.
Built Bot Analyzer — an LLM-powered configuration analysis feature on Amazon Bedrock and Step Functions, live across eleven AWS regions. Benchmarked twelve LLMs for accuracy, latency, and cost. Also designed Multi-modal Conversation for SMS capture during live voice calls.
Led region build automation cutting cycle time from fourteen to five days. Co-designed a multi-dimensional token-bucket throttling system — load-tested at 7× peak production traffic — protecting millions of Lambda users from cross-tenant spikes.
"I ship the thing.
Measure the thing.
Improve the thing."
— operating principle
LLM-driven configuration analysis for Amazon Lex chatbots, surfacing alignment with AWS best practices through public API and the Lex console. End-to-end across 11 regions; 12 LLMs benchmarked for accuracy, latency, and cost.
SMS input capture during live IVR voice calls without interrupting the session. Extended bot channel APIs to integrate End User Messaging with Amazon Connect. Delivered as a private beta across four enterprise teams.
Led automation of Lambda's EventBridge control-plane region builds — replacing a fully manual sequence with an orchestrated CI/CD pipeline. Cycle time from 14 days to 5; zero manual steps and zero errors.
Co-designed a multi-dimensional token-bucket throttling system protecting the Lambda fleet from cross-tenant traffic spikes. Per-host local cache; async sync to data plane. Load-tested at 7× normal traffic; p99 overhead under 25 ms.
— Han Cao
