At BOSS Zhipin, I replaced rule-based targeting for high-priced recruiter packages with uplift-model recommendations. I fused revenue and paying-rate uplift models, using parameter search, log replay, and A/B tests to shape the scoring; the fully launched strategy increased ARPU by 2.1%.
I also found that candidates who had previously used free chat were permanently filtered out of search, and changed the window to 90 days. The fully launched change raised paying rate by 6.48% and ARPU by 5.68%. To improve measurement, I worked with product and backend to connect conversions to search requests and built a reusable full-funnel evaluation table.
At BOSS Zhipin, I independently built a team data-science agent on the open-source DeepSeek Harness with the company-hosted DeepSeek model. It supports recurring analyses and team Q&A, with human confirmation before full data pulls and review of proposed knowledge updates. Earlier, at Kuaishou, I built a pandas tool to separate retention changes into within-group and mix effects, helping identify low-quality channels for shutdown.
