KAIWEN ZHU

• Designed and trained a multi-task regression model using PyTorch and TensorFlow to predict video engagement metrics (click-through rate, like rate, watch completion rate) from content-level features.
• Utilized MS Excel to gather and summarize weekly video data, supporting video publishing initiatives.
• Created graphs, including histograms and pie charts, using Python and MS Excel to visualize data, driving customer comprehension.
• Conducted data analysis on large datasets; compared and interpreted data generated week over week to provide recommendations to clients.
• Collected and sorted data from multiple outputs of a manually annotated model; trained a reward model for predicting user preference in model outputs.
• Analyzed relevant data and verified its accuracy while removing irrelevant data points; summarized and reported data to provide information that supported business objectives.

• Engineered TA/ATA Training Appointment/Notification System to automate lecture scheduling and enhance training efficiency.
• Streamlined homework submission process by creating automated system for saving and distributing sales scripts to Teams groups.
• Designed notification feature for posting tailored updates in specific Teams groups based on time and content.
• Built scoring system for Tesla’s 2 and 3 Stars Sales Certification programs to enable accurate trainee performance evaluation.
• Utilized Power BI to develop registration system dashboard and automated Python scripts for generating weekly reports.

This project focuses on analyzing movie reviews using machine learning. It includes three tasks:

• Sentiment Classification — Determine whether a review is positive or negative.
• Score Regression — Predict a numerical score (0–10) associated with a review.
• Critic Identification — Identify whether a review was written by a top critic or a regular user.

We explore two approaches:

• A TF-IDF + traditional ML pipeline for benchmarking.
• A BERT + tree-based model pipeline for contextual, high-performance prediction.
• BERT-based models outperform TF-IDF baselines in sentiment and score prediction, offering better generalization and reduced overfitting.
• Score regression using BERT and XGBoost achieved strong alignment with ground truth scores.
• Top critic classification remained challenging due to class imbalance, though BERT improved minority class performance.