I'm Gordon Tsai.

GPA: 3.84 / 4.0

GPA: 3.78 / 4.0

• 1. Tasked with enhancing GEICO's chatbot efficiency to reduce response time and improve customer service by addressing existing system limitations.
• 2. Built the development of advanced NLP algorithms using Transformers and LSTM models, and fine-tuned GPT-2, to accurately understand and categorize customer intent.
• 3. Containerized chatbot using Docker, implemented CI/CD pipelines with GitHub, and deployed solutions to cloud providers (e.g., AWS), ensuring consistent and scalable deployments.

• 1. Cleaned, transformed, and merged 100,000+rows of data and implemented regression model and feature engineering from scratch for the fleet management optimization of electric trucks.
• 2. Used principal component analysis (PCA) to reduce the feature space, and leverage multiple linear regression to forecast the battery level for truck and using logistic regression to predict the probability of brake failure.
• 3. Optimized Fleet Management Using battery level forecasts and warning level prediction to reduce the operation.

• 1. Started with the item-based collaborative filtering for movie recommendation, which is based on KNN and Cosine similarity for nearest neighbor search to avoid “Curse of Dimensionality”.
• 2. Improved the movie recommender by utilizing Alternating Least Square (ALS) Matrix Factorization to overcome the shortcoming of popularity bias and item cold-start problem.
• 3. Conducted model hyper-parameters tuning with ML cross-evaluation toolbox and monitored data processing performance via Python.

• 1. Started with the item-based collaborative filtering for movie recommendation, which is based on KNN and Cosine similarity for nearest neighbor search to avoid “Curse of Dimensionality”.
• 2. Improved the movie recommender by utilizing Alternating Least Square (ALS) Matrix Factorization to overcome the shortcoming of popularity bias and item cold-start problem.
• 3. Conducted model hyper-parameters tuning with ML cross-evaluation toolbox and monitored data processing performance via Python.

• 1. OLS we will report confidence intervals and heteroscedasticity robust standard errors for our treatment variable and all other pre-treatment variables. We will also report the p-value for the treatment variable.
• 2. Use Double Lasso and Elastic Net, we will report confidence intervals and standard errors. For Double Lasso, we will report which variables had non zero coefficients.
• 3. Find out that higher dropout rates could be correlated with lower income areas where financial education would make a bigger difference.

• 1. Built Fama-French 3 factors model using size of firms, book-to-market values, and market premium to predict S&P 500 stock return.
• 2. Backtest strategy performance by constructing a monthly rebalanced long only portfolio.
• 3. Test the traditional momentum model as an additive factor on the Fama-French model.
• 4. Using CAPM model to construct an alternative momentum factor to isolate beta effect and capture idiosyncratic stock performance.
• 5. Rerun back-testing and Sharpe Ratio increased from 1.04 to 1.28, with significantly lower vol and drawdown.
• 6. Applied Long Short Term Memory (LSTM) Model to predict the relative performance of each sector based on the historical pricing data.
• 7. Further improve the strategy by taking the sector allocation effect into consideration.

• 1. Performed imbalance data cleaning and feature selection on origination and recurring factors, including delinquency, FICO score, and DTI ratio, etc. to determine the main causes of the default activities.
• 2. Constructed a variety of machine learning models, such as Logistic Regression, Decision Tree, Random Forest, Adaboost, SVM, and Neural Networks, to forecast loan default using Python (Sklearn).
• 3. Applied cross-validation to ensure model robustness and limit the risk of over-fitting, and compared model performance based on the ROC curve and AUC. The model has an 89% accuracy in the out-of-sample test.

• 1. Applied causal observational study to discover the causal relationship between the number of unique donations a presidential candidate received and the proportion of the vote they received in the primary.
• 2. Constructed Causal Directed Acyclic Graphs (DAG) to reveal the causal relationship between the treatment and the outcome, and used the Pearl’s Back-Door criterion to determine the minimum set of confounding factors.
• 3. Performed OLS regression adjustment, inverse probability weighting and propensity score matching to tear out the effect of confounding factors, and estimated the causal impact of number of donations on proportion of vote.