I founded AhaaaMath(http://www.ahaaamath.com/en), a startup company, in 2018.4. AhaaaMath is a technology-driven education company dedicated to promoting artificial intelligence, visualization, and gaming technology in the educational applications. The first goal of the company is to build the best educational technology and products for elementary mathematics education.

As the director of Ads Quality Center , Department of Social and Performance Ads, Tencent, I led a team of 60+ AI engineers, in charge of ads quality optimization. My team strived to accurately understand Tencent users and deliver right ads to the right users at the right time，tradeoff the benefits between ads platform and advertisers, and ensure reasonable user experience as well. The main responsibility of my team was to improve advertising efficiency through large-scale machine learning, including user data mining for audience targeting, click-through rate prediction, conversion optimization, as well as user ad experience optimization .

– Led the AdQuality team of 18 software engineers participated in the development of a contextual advertising system similar to Google’s AdSense. Our team is responsible for the quality of the query-ad matching and the final ranking of the candidate ads. Our work includes keyword extraction, text classification and clustering, large scale topic modeling, user behavioral targeting, and click-through rate prediction of ads. All the components involve the techniques in natural language processing, parallel machine learning, and big data processing. As a team leader, I participated in the design of all the above components, as well as the implementation of some key components. – Ported Google’s code review system Rietveld into Tencent, and participated in the establishment of the code review culture in our division.– Participated in the design and development of the A/B test experimental platform in our division.

– Mapreduce-Lite, a light MapReduce C++ implementation which is mainly used to support the development of parallel machine learning tools in Tencent. The API interfaces are similar to the one used in Google, and we implemented several parallel machine learning tools based on it. The tool was finally released as an open source project (http://code.google.com/p/mapreduce-lite).– Parallel Logistic Regression (with L1 regularization), two training algorithms were implemented in this parallel tool, including OWLQN (a variant of LBFGS) and SGD. We cooperated with the sponsored search team to do feature selection and ads CTR (click through rate) model training, and our model improved the ads CTR by 19%.– Parallel LDA(Latent Dirichlet Allocation), a latent topic analysis tool for text modeling. Our algorithm is much better than the standard LDA algorithm in two aspects: we improve the training speed by 40 times, and the optimal number of topics is automatically trained rather than specified by human. The trained topic model was used widely in our products, including query expansion, ads indexing and retrieval, query-ads relevance computation; also we applied the topics as features in our text classifier and CTR model, which significantly improved the final result. - Participated in the click modeling project, used machine learning approaches to model users’ click feedback of our web search engine in order to improve the ranking quality of webpage. The click model used here is the Yahoo’s DBN (Dynamic Bayesian Network) model. My duty was to design the feature generation framework and implement different features. The project significantly improved NDCG relevance improvement.- Participated in developing natural language processing software, including statistical Chinese word segmenter, Chinese new word identification, name entity recognition, entity words mining from search query log and web news corpora.

- Associated a graduate level course “Java Programming”

- Associated a graduate level course “Introduction to Algorithm”- Associated an under-graduate level course “Probability and Mathematical Statistics”

Participated in the Internet Documentation Project