Live Streams Recommendation, Graph Embedding:
Enhanced internal ML trainer (C++) performance by mitigating communication overhead. Reduced mini-batch forward latency by 40%+ in certain graph embedding training circumstances.
Extracted relations from Petabyte log data and built distributed user-author graphs using MapReduce. Devised graph encoders and trained an embedding of graph nodes using Tensorﬂow and ByteDance ML API.
Integrated end-to-end embedding to Click-Through-Rate prediction. Increased online user staytime (+3.5%), comment rate (+3.8%), and other metrics signiﬁcantly in AB tests for Douyin (TikTok for Chinese market).
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. Decreased the usual day-long response time to <1hr in a recent dataﬂow accident.
Constructed a pipeline that analyzes importance of 300+ features in a Monte Carlo fashion and performs feature modiﬁcation on Terabyte model checkpoints on clusters based on results. Saved 2hr+ manual labor per iteration and 35k+ core-hour computing resources in total than hand-tuning.