Experience

Work with AI Foundation team, building AI Model Quality Assurance Solution for all LinkedIn AI products.
Machine Learning Model Monitoring: Core algorithm innovator and developer of the health assurance anomaly detection system. The algorithm/system plays a central role in ensuring LinkedIn's AI model healthiness.
AI Model Understanding: Core developer of a central model understanding and interpretation library for instance-level and global-level feature importance calculation.

Smule

Batched Recommendation: Completed adaptive music recommendation systems that recommend trillions songs and millions users. Administered and optimized various recommendation algorithms behind Explore and Songbook pages.
Real-time Recommendation: Improved web user engagement by 10% by building as a chief developer a brand-new real-time music recommendation system. Implemented modules including real-time feature extraction, ranking, database communication, and training modules. Used tools including Hadoop, Scala, Spark, Flink, Cassandra, Redis, Kafka etc.
Features Extraction: Improved users’ average music listening time by 30%, 12%, and 5% separately by developing three new features for personalized recommendation models behind Feed, Explore, and Songbook pages. (co-authored in publication).
Intention Extractor (NLP Research): Led researches on text classification and keywords extraction, trained model on comments data with tools TensorFlow and NTLK.
Other: Learn new techniques in industrial recommendation system through paper, conference, or Coursera. Apply the new techniques in to Smule's machine learning model.

Anomaly Detection: Largely reduced company’s costs in data labeling by inventing a semi-supervised anomaly detection algorithm used for detecting vehicle’s problematic sensors while still keeping high accuracies. (co-authored in publication).
Data Visualization: Facilitated engineers’ diagnosis work by developing a ML-based visualization tool, leading to 20% time saving.

Mental Illness Predictor: Pioneered in predicting people’s mental-health status from social-media posts, self-motivated from data collection through multi-model feature extraction, prototyping, and model performance testing. Leading to 80% F1-score in prediction.

Language Detector: A French/English/Italian language detection classifier using neural network trained on public dataset.
3D Reconstruction: A deep convolution neural network for 3D re-construction from 2D images, implemented with TensorFlow.
Music Highlight Extractor: A music POI extractor based on historical engagement data, implemented with Spark.