Instagram Ads Delivery and Ranking

• present a new natural language processing approach to time series forecasting by using a combination of bilinear projection and self-attention mechanism to learn complex patterns and dynamic information from time series data

• develop a generalized language model that could be used for univariate and multivariate time series data across various possible domains that are not just language data, specifically focusing on financial and astronomical time series data

• propose a possible pipeline of confidence measures to improve the usefulness of the language model by deriving confidence interval using Bayesian inference

• Redesigned Microsoft O365 cloud backup storage layer to provision multiple storage accounts for big customers, and rescaled cluster from supporting 30k users to 500k users

• Implemented load-balancing strategy for Azure storage backup users, improving storage limit from 550 TB to 200k TB

• Designed storage sharding product design document in a team of 6 people, and collaborated across two teams

• Skills used: Go, SQL, MySQL database, Microsoft Azure, grpc

• Investigated interaction and gaze patterns of 65 students on the spectrum engaging in an existing, science-oriented education technology using Python

• Collected a video Q&A dataset with 400 social content and developed related research questions

• Developed a multi-agent computational cognitive model of theory of mind reasoning as the primary data researcher and team leader; used PySWIP to write the production system built on the model

• Designed an education application that provides visual scaffolds for students to improve their theory of mind reasoning skills using gamification and user-centered software design in a team of 5 people; successfully tested on 150 middle school students