- Applied OpenAI's Whisper audio-to-text model to transcribe customer-agent call logs, constructing a word-level matching algorithm for speaker diarization to improve speaker segmentation and extraction of withdrawal reasons from loan applications

-Utilized CatBoost model with hyperparameter tuning for feature selection among over 200 momentum and directional indicators for Forex trading, identifying the top 20 most important indicators for further development of indicator formulas- Constructed a mathematical strategy for developing indicator formulas based on the historical distribution of directional indicators, which track the continuous positional relationship between two price-related indicators, such as SMA indicators, leveraging min-max scaling, piecewise fitting, and backtesting with TradingView

- Implemented Meta Segment Anything computer vision model to efficiently isolate 2,670 buildings from 2,910 thermal images, and conducted a quantitative analysis to detect potential leakages in the building images, providing 73 buildings with suspected leakages to engineering team for further physical inspection- Applied the ESIM pre-trained NLP model for building address matching, increasing matched addresses from 5,000 to 22,000 out of 52,000, significantly enhancing data precision compared to the previous method- Benchmarked air conditioner identification in Google Street View images using the LangSAM model, achieving 84% accuracy and 88% recall on a dataset of 274 buildings; transitioned to AWS SageMaker’s object detection model with ResNet layers for enhanced feature extraction, fine-tuning it on a self-masked dataset to achieve 72% accuracy on a larger dataset of 2,000 buildings

- Hosted 4-hour weekly Office Hours as a Teacher Assistant for CSOR 4231 (Analysis of Algorithms I) at Columbia University, clarifying key algorithms BFS and DFS, elucidating Ford-Fulkerson, and demystifying NP-completeness, bolstering students' algorithmic understanding and course performance

- Conducted a user preference survey for ticket icon design leveraging A/B testing methodology and analyzed the result with R to identify key factors influencing users' preferences, contributing to a remarkable 17% increase in user click rates for the ticket icon- Employed SQL to derived historical sales data from the database, predicted the Gross Merchandise Volume of summer sales by developing an ARIMA model based on historical data for holiday sales, facilitating data-driven decision-making for strategic planning- Leveraged logistic regression, Random Forest, Naive Bayes, and XGBoost algorithms to predict potential users, driving 77% model accuracy and 82% recall, resulting in a 5% accelerated growth of active users in the Ticket Department