o Built a python toolkit for model prediction difference analysis and performance analysiso Presented the results to SVP, SLTs, and team members

o Developing web-scraping tools to collect news articles and perform semantic analysis to facilitate automated ESG analytics.o Developed algorithms for Event Extraction and Knowledge Graph construction regarding geopolitical-important entities.

o Participated as part of Penn NLP Lab team member in the Amazon Taskbot Challenge and was responsible for document reranking module development and testing.o Incorporated SentenceBERT reranking algorithms and improved MRR by 10%; the team won the 4th place.

o Initiated and developed a Batch-Testing Toolbox for the team to replace manual work with automation:- Regression Tester: a C# application that automates manual regression tests and generates detailed reports, which reduced testing time by 80% and eliminated potential operational risk in manual testing.- Test Case Generator: a data ETL program built with HiveQL, PySpark, and C# that automatically generates test cases for the production data pipeline, increased test coverage by more than ten times.- Variable Dependency Parser: a Python program that validates feature derivation documentation by checking dependency conflict and performs a topological sort to eliminate re-calculation. Increased model inference speed by 50%.o Deployed, tested and debugged a Customer Complaints Prediction Model with the Experian Strategy Design Studio (SDS).o Gained substantial experience in consumer finance business model; learned the ML models supporting various parts of the business.

Design and developed a prototype Federated Question Answering System over Knowledge Graph

o Participated in the project to enhance FICO Score performance:- Ensembled Scorecard with XGBoost algorithm to explore interaction relationship between predictors, which increased the overall model performance by 12% (5% higher KS and 2% higher AUC).- Reduced model complexity by 75% using Elastic-net Penalty (mix of L1 and L2) and a novel Weight of Evidence Categorical Data Encoding Method, which improved model interpretability and increased robustness against overfitting.- Implemented an R library and hosted a training session covering model development, tuning, and performance evaluation.o Optimized and upgraded legacy R data ETL pipeline to automatically report missing rates, potential outliers, feature stability (PSI), and feature predictive power (WOE/IV, KS, and AUC), reducing EDA time consumption by 60%.

- Initial Mortality Minder project for the AHRQ Data Visualization Challenge- Worked as the backend developer and manager in developing a web-app to analyze and present the rising trend of county-level abnormal mortality and potential socioeconomic causes in the U.S.- Designed the general app architecture, including the main control logic, data flow, and package management.- Conducted data wrangling and visualization based on tidyverse package set, and development using RShiny framework.- Applied K-Means clustering algorithm to group counties based on mortality dynamics pattern, and applied the Kendall Correlation Analysis to investigate relationships between mortality and socioeconomic statistics.- Provided the 5-member backend team with technical training in R programming and assigned tasks based on team members’ academic background & programming skills.- Delivered presentation of the app to industrial leaders (e.g., VP of CDPHP) and scholars on behalf of the entire app development group