– Large-scale Ads pipeline and featurization development∗ Built large data pipeline and aggregation featurization pipeline for user, Ads, query and other features,so that the system could capture long term user behavior and click through rate (CTR) drift∗ Significantly improved model performance which increased about 1.8% revenue for a market of 2 billionwhile reduced system latency and memory cost by 60 percent∗ Further developed features using search result data and real time data that updated in 40 seconds onlineto enable real time feedback of Ads system∗ Drive the pipeline updates to keep its long term performance for a team of five scientists– Ads decoration modeling∗ Drove modeling work for a team of four to enable 8 new Ad decoration types∗ Built first Ads decoration model for our partner traffic using a gradient boosting tree model that boostedon top of a neural net, which has 3 million direct revenue impact∗ Applied feature selection method to keep the model performance with 30% memory cost– Causality modeling in click and conversion rate models∗ Trained gradient boosting tree using exploration data to model Ads layout as confounding variable∗ Predicted normalized click through rate (CTR) which corrected CTR bias caused by layouts∗ Generated features using the normalized CTR that guarantee long term stability, solved the problem thatmodel converges to show only small portion of Ad candidates∗ Applied DNN model to decouple the causal factors and predict conversion rate (CVR) to enable auctionadjustment for hundreds of verticals, otherwise these traffic won’t be efficiently monetizable– Deep neural network and semantic features∗ Delivered a deep neural network model for semantic and id features that is scalable to new advertisingpartners∗ Applied DNN model to generate query and Ad embedding vectors∗ Enabled the DNN model for online learning, drove the the online learning maintenance with 3 SDEs

– Click through rate modeling∗ Merged models and data pipelines for 200 partners in a large-scale Ads system and consolidated datasampling and clean-up for data of about 0.9 billion per training∗ Developed and iterated machine learning models for Ads recommendation system serving about 10 million dollar revenue per day– Featurization of Ad partners and investigated model score parity∗ Designed and implemented features for Ad partners and its interaction with query and Ad∗ Designed an online-offline machine learning model score parity module with visualization that has beenused in our team of 40 data scientists, saving hundreds of hours for debugging per year

 - Taught Calculus I, II and lab session of Calculus III in classes of size 30 or 120 - Gave lectures, hold office hours, assigned and graded exams and homework - Instructed 120 students to learn mathematical software, such as Matlab and Mathematica

 - Discovered new techniques to control distributed time-delay system using spectrum design - Experimented and improved the Lambert W function numerical method in Matlab - Methods applicable to cutting edge engineering projects such as autonomous driving