At Alibaba International, I built a 300-case gold benchmark from historical merchant root-cause analyses and calibrated an LLM judge against human ratings, reaching 92% agreement.
I mined agent traces and tool calls for regression cases, helping raise offline pass rates across two workflows from about 71% to 84%. I also collected and filtered teacher trajectories for action-only SFT and designed verifiable rewards for online GRPO; strict task success on held-out shopping tasks rose from about 53% to 67%.
At Kuaishou, I built a stratified text-to-SQL evaluation set from 260 real analytics questions and added data-quality checks. Correct table selection rose from 75% to 88%, while SQL execution success increased from 72% to 86%.
For an independent DuckDB project, I built a data pipeline repair benchmark using NYC taxi-trip data across two monthly Parquet releases. Its automated verifier passed the reference SQL repair and detected failures involving dropped rows, NULL handling, and mismatched monthly fee totals.

