Co-led development of Microsoft Dynamics Copilot features leveraging LLMs and RAG (retrieval augmented generation)

• Collaborated with teammates to develop two NLP models -- Summarization model (BART model) and Explanation model (BERT-SpanKPE model) as new features in Dynamics 365 Omnichannel, with the goals of increasing agents’ productivity and providing customers with a better customer service experience.• Implemented two HitApps with JavaScript and launched them on UHRS to retrieve human evaluations for these two models to ensure quality outputs.• Preprocessed the training data and contributed to improving the Summarization model performance by 8.75% by parameter tuning, feature selection etc.• Benchmarked the Explanation model with other SOTA keyphrase extraction model and assessed our model’s lift using Precision, Recall and F1 score as evaluation metrics.• Participated in Microsoft’s worldwide Hackathon and collaborated with teammates to implement a smart reply engine which automatically generates replies and confidence scores of them regarding the conversation.

• Collaborated with engineers to develop with Tensorflow the recommender system’s ranking model using the Wide & Deep learning framework for Kandian Shipin, a short video distribution application. • Coded the LightGBM with Python for feature selection, extracting useful features for the ranking model from over 100 candidate features, categorized by user, item and context related.• User average view time improved by 3 percent after the top features selected by lightGBM are added to the ranking model. • Launched the online evaluation based on Time Dimension and the A/B test that accounts for algorithmic causes to gather user responses for the recommender system developed.

• Optimized the regular feature selection method by using Frequent Pattern Mining, to reduce the number of features and to obtain higher accuracy than baseline measures using the existing feature selection methods.• Coded a program that applied the Frequent Pattern Tree (FP-Tree) to find sets of features that appear frequently among the top features selected by the main feature selection methods.• Conducted experimental evaluations with two datasets containing small and very large number of features using cross-validation to build models and obtained roughly an improvement of 10.2% in predictive accuracy compared to the baseline measures.• Paper in submission to the Association for the Advancement of Artificial Intelligence (AAAI).

• Implemented with Hadoop and Python a program to obtain site names from tag titles.• Utilized Git and Hadoop to extract user history information from company database including URL, HTML tag title, visited frequency, etc. • Developed with python an application following the decision tree algorithm to analyze the fragments of tag title appearing in different domains and webpage source code.• Applied the application to get potential site name candidates.