Microsoft AI Platform- Nebula: fast checkpoint save 1000x than torch save.- AIP Data analysis

Microsoft - Bing1. Built up data-driven model to promote user into content feed view on Microsoft Edge new tab page (NTP)2. Drove and built the WPO feature system to provide daily and Near Real Time features for WPO model training.

Microsoft - Bing• Built up keyword extraction pipeline in financial news. Utilized AutoPhrase to mine dominant phrases, TF-IDF to extract keywords and combined NER to identify the entities in news. Experiments results show good accuracy and receive positive feedback from the judges.• Be responsible for the research and implementation of similarity comparison algorithm in sentence level. The idea of QA Similarity was adopted in this project, which contains three layers: Siamese network learning layer, ANN index layer and ESIM re-sieving layer, speeding up the calculation process while ensuring high precision similarity comparison.• Refactored backed code for NLP Group to simplify the process of updating for new models.