- trained models to predict delivery time window for last mile delivery- trained models to understand customer intents and recommend products on Amazon.com- developed chatbot and agents using LLM

 Lead to design, implement and ship an end-to-end entity linking pipeline that extracts key local entities from free-text articles, supporting multiple Bing local search scenarios about local businesses and places. Lead to design, implement multiple features to improve Bing local product query relevance, and shipped to global markets. Delivered a multi-lingual product intent deep learning model, a more resilient query relevance scoring model, a fair and accurate query relevance evaluation model, and a rule-based query results ranker. Designed and implemented a training pipeline of a named entity recognition model using BERT as a key component to enable the enterprise search for enterprise clients, bridging the gap of confidentiality and beyond-the-compliance of the enterprise data. Designed and implemented a query component for querying knowledge graphs with natural languages, enabled a much more natural and smooth interaction between enterprise clients and enterprise knowledge graphs. (Filed 3 U.S. patents and published 1 conference paper)

 Worked on Streaming Data Characterization project (https://aim.pnnl.gov/projects/streaming-data-characterization) Enriched and formalized the concept of semantic importance that models the streaming data importance. Designed and implemented a generic framework in Java to process a wide range of streaming and static inputs. Demonstrated the superiority of both the concept and the framework in the streaming context: together they saved 253.71x more memory, delivered 7.6x faster response time, 2.38x more precision, and 1.86x higher throughput. Presented 2 talks and 1 poster on the project deliverables, and received high evaluations from colleagues.

 Worked on Analysis in Motion Initiative (https://aim.pnnl.gov/) Designed and implemented a scalable stream reasoning architecture in Java using modern graph databases. Proposed an innovative notion called semantic importance that describes the streaming data importance. Implemented 2 use cases in both sport and cybersecurity domains to validate the hypothesis that the semantic importance can enable smart and flexible window management strategies in stream reasoning settings. Quantified the semantic importance benefits as high precision (90%), and high recall (97%) in short time (0.81s). Published and presented 3 first-authored conference workshop papers at ISWC, the best semantic web conference.

 Project mentor for a group of undergraduates who work during the summer on a visualization project  This project aims to visualize a complex network of drugs, proteins and diseases with 12 projectors and on a huge 360 degree, about 4 meters high projection screen.  Provided both technical and design guidance for summer working students. Organized the final deliverable presentation

 Recognized by the award of Who's Who Among Students in American Universities and Colleges in RPI. Steered CSSA administration reformation to reduce 10 redundant departments into only 4 main ones. Mentored department heads to organize 1 basketball tournament, 1 esport tournament, and 2 celebration parties. Budgeted fundings with Rensselaer Union, fundraised with outside sponsors, planned about $5000 for one year. Represented CSSA to advocate Chinese culture and celebrate the diverse international culture at RPI.

 Developed a proof of concept semantic application on better medical devices management Developed a simple algorithm to convert unstructured text files into structured RDF files leveraging NLP techniques Developed both front end UI and back end services to provide specific contents delivered properly and smartly  Exploited semantic technologies such as RDF, NLP, graph database(AllegroGraph) etc in the project.