Amazon Ads

Self-service Machine Learning / Feature Engineering Platform under Core Services* Online Feature Computation: Streaming data is ingested from SNS, Kinesis, or internal data stream services, and the compute logic is executed on our fleet. We have our own proprietary streaming process engine.* Offline Feature Computation: Batch data comes through S3 (or other internal data storage), triggering a Spark job that runs custom logic (which shares the same logic as the online process). Spark is wrapped within our platform, and customers won't have direct access to it. The results are written to our internal storage.* Feature Retrieval: We have a dedicated service with 20,000+ hosts to support more than 50 clients with 3 million TPS during peak times.* Pluggable: All components, including feature data sources, feature compute platforms, and feature storage, can be replaced by customers, allowing them to customize their data streamline processing.* Tools: We have an automated code analyzer to help customers follow best practices. We also provide a user interface that offers control over every feature, supporting real-time retrieval testing, read/write metrics, audit logic, and dependency lineage through metadata analysis.

- Velocity Compliance Threshold - Design and lead teams for Velocity Compliance Limit feature by by leveraging Linear Regression to avoid 1000+ customers per week blocking by disbursement hard limit's KYC check.- Manually Risk Override - Design and lead teams for MRO which allows investigators to manually change Risk Evaluation Machine Learning Model output for particular sellers.

* For Carmacam project. http://www.carma-cam.com/* Implemented a real-time (30+ fps) DUI (Driving Under the Influence) detection system by leveraging modern deep neural networks (R-CNN) and traditional computer vision algorithms (canny edge detection / KNN).

* Intern in Core Machine Learning Services(CTPS). Got the return offer from Amazon.* Designed and developed (both frontend and backend) a web platform for feature engineering, which enables data scientists to maintenance pipelines without coding efficiently. The final platform yields a 3-fold improvement in the person-hours required to launch new pipelines and find at least two existing bugs in the system.

◦ Stackable Attention Module: Designed a stackable self-attention module using rank-pooling (Based on the idea of Learning to Rank) to extract attention map from the temporal dimension in ResNet for video recognition. Submitted to IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI, 2018 Nov).◦ Online Rank-Pooling Algorithm: Designed a real-time online algorithm for calculating Rank-Pooling in video streams. Giving a mathematical proving for the equation between the original and our novel algorithm.◦ 3DCNN Baseline Model: Designed a Global-Local Focus 3DCNN model as the new baseline of the ChaLearn LAP Large-scale Isolated & Continuous Gesture Recognition Challenge. Performance of the new model surpasses all the existing model in the competition. Submitted to IEEE Transactions on Multimedia (TMM 2018 Oct).◦ 3D Spacial Transfer Network: Extend 2D Spacial Transfer Network to 3D. Using depth camera and RGB camera to generate 3D point cloud and training a network to find the best direction for gesture recognition.◦ 1x1 Convolutional Fusion Structure: Designed an adaptive weight fusion scheme using 1x1 Convolutional kernel, for ensembling features in Gesture Recognition. Accepted by 13th IEEE Conference on Automatic Face and Gesture Recognition (FG 2018) as an oral paper and Machine Vision and Applications (MVAP 2017).◦ Labeling Tools: Implement a Human Facial Attribute Labeling Tools of Facial Multi-Variate JFA Dataset. Submitted to IEEE Transactions on Cybernetics (TCYB 2018 Mar).◦ Evaluate Competition: Re-implement and evaluate competitors’ models in different frameworks (Karas, Tensorflow, Caffe, Pytorch) for the ChaLearn LAP Large-scale Isolated & Continuous Gesture Recognition Challenge. Accepted by International Conference on Computer Vision (ICCV 2017 Workshop).

• Designed the tokenizer for semi-structured log data of logs analysis framework LANE.• Analyzed large scale (1TB of new log data per day) transactional logs of Microsoft Azure and Office 365 on Cosmos to detect potential cloud system failures and help the developer to focus on high-impact bugs.• Developed an adaptive linkage strategy for hierarchical clustering to increase the accuracy of log message clustering on Cosmos.• Developed a new distributed histogram tree algorithm, which differs from the old synopsis algorithm, to speed up the query in high-dimension data for IN4 (PowerBI) Project.

• Developed an efficient spider to collect and update all the information of software company in Macau automatically.• Participated in designing and implementing a human resources management web-system.