• Worked at AlternativeUtteranceService (AUS) to improve Alexa conversation user experience, by providing better utterance rewrites, with state-of-the-art AI technology.• Developed the first Deep Learning based utterance rewrite pipeline, contributed million level good rewrites traffic increase weekly and 7% CPDR (Customer Perceived Defect Rate) decrease.• Led the engineering design and development for Alexa personalized rewrite pipeline with multi-source customer history index.

• Worked at Amazon new internal accounting analysis platform. Designed and developed anomaly detection module for Amazon transaction data.• Identified risks and anomalies from Amazon bank account activities by empirical and unsupervised methods. Saved more than 40 accountant hours of verification time for the accountant team every month.

DentistAI diagnosis of dental conditions• Conducted dental images cleaning, affine transformation, augmentation• Built and optimized state-of-the-art CNN models for the X-ray image classification on both picture and pixel-wise

Multilingual Question Answering on BMW car manual data• Trained state-of-the-art attention models with multilingual word embeddings.• Aligned multilingual data by the machine translation model and multi-version car manual's data.• Built novel knowledge discovery ensemble models for single language QA.

• Developed an ensemble Stacking package for internal machine learning platform in scikit-learn style.• Deployed the package into practical classification tasks.

Video Recommendation System based on Video Affective Analysis• Extracted nineteen video features, using Support Vector Machine to map these features onto the Lövheim cube.• Applied Grey Relational Analysis (GRA) to video affective analysis; Proposed a novel recommendation algorithm with strong scalability.• Developed a video recommendation website for experiments; evaluated the performance improvement of recommendation accuracy with both parameter metrics (Precision, Recall, F1 and RMSE) and user study (A/B test, User Satisfaction).Music Mood Classification based on Signal Analysis• Implemented the orthogonal matching pursuit algorithm with Gabor functions in Matlab for music mood classification.• Extracted sparse coefficient vector of given music signal and uses it to construct eleven time-frequency features.• Compared with MIREX 2011 Champion method, resulting in a 1.26% increase in single-label classification accuracy.Partitioning Algorithm Research for Mobile Cloud Computing Offloading• Designed a Max-Cut Graph algorithm to solve the optimistic partitioning problem.• Proposed a cost evaluation model called CMET (CPU-Memory-Energy-Time) to compute the weights of edges of the Max-Cut Graph.• Assisted with conducting a survey on current offloading schemes such as MAUI, CloneCloud, and MapReduce.