Live Streams Recommendation, Graph Embedding:

∗ Improved the performance of the internal ML trainer (C++) by reducing communication overhead. Reduced the mini-batch forward latency by over 100% in certain training circumstances that involve graph embedding.

∗ Extracted relations from Petabyte log data to build distributed user-author graphs using MapReduce and designed graph encoders to train an embedding of graph nodes using Tensorflow and ByteDance ML API.

∗ Integrated end-to-end embedding to CTR prediction and significantly increased online user staytime (+3.5%), comment rate (+3.8%), and other metrics in AB tests for Douyin and TikTok.

Systems for Engineering Efficiency:

∗ Designed and developed a system from scratch in Django with RESTful APIs that creates and manages alerts for 100+ online models of 5 products and presents model health status on a dashboard. Attracted 70+ internal users and developers. Reduced the usual day-long response time to less than an hour in a recent dataflow accident.

∗ Developed a pipeline that analyzes the importance of 300+ features in a Monte Carlo fashion and performs feature modification on Terabyte model checkpoints on clusters based on the results. Saved 2hr+ manual labor per iteration and 35k+ core-hour computing resources in total than hand-tuning.

Baggage Re-Identification for Smart Security Inspection: Extensively investigated past architectures for person and vehicle re-identification (re-ID) tasks. Implemented multiple architectures and evaluated their performance for baggage re-ID. Achieved a 0.76 accuracy of CMC rank-1 (improving the baseline by a 0.34 margin in accuracy and about 100x in speed) on the overall re-ID task (image retrieval from the 500-baggage gallery) on the Multi-View Baggage dataset.

Testing Framework Development: Developed a fork upon the open-source deep-person-reid framework for the team’s future research and deployment workflow, including features such as activation visualization, training checkpoint management, Comet.ml integration, and CLI tools.

Instructor of Theory of Computation: Co-taught the 5-day seminar on theoretical computer science, covering topics such as discrete math fundamentals, computability (DFA and TM), and efficiency (polynomial reduction and P vs. NP). Designed slides, handouts, and assignments.

TA for AI Courses: TA-ed 10-day main courses. Designed homework and assessments for neural net basics, Markov Decision Process, Q-learning in Deep Reinforcement Learning (2018), and feature extraction, SVM, DP & Seam Carving in Computer Vision (2019). Led Q&A sessions and graded homework.

Academy Experience Design: Directed the Academic Team of TechX 2019. Led the processes of course design, staff search, and infrastructure setup for 5 courses.

Product Management: Assisted the development of the platform from scratch using TypeScript and Node.js; generated user experience reports and feature addition proposals; beta-tested the platform with 100 users.

Business Operations: Drafted business plans and pitched to Tsinghua alumni investors, ZhenFund, and various other VCs; Invited 15+ Chinese corporations and organizations to enter Fintern.