“Influencer Selection with MMOE” with Prof. Shunyuan Zhang from Harvard Business School and Prof. Xitong Li from HEC, Paris.

• Improved Multi-Gate Mixture-of-Experts (MMOE) algorithm, based on the incentive mechanism of influencer behavior and interactions between marketing goals (reputation, revenue, acceptance), increasing the revenue prediction performance by over 20%.
• Explored the underlying mechanisms of multi-task learning models, based on literatures and regression analysis, identifying model-relevant factors such as the gating mechanism, data-relevant factors such as data sparsity, and feature-relevant factors such as latent relevance.
• Embedded dynamic weighting algorithm (GradNorm) into MMOE to dynamically tune the weights of different tasks during the training process, increasing the model performance by over 200% over a highly unbalanced and sparse dataset.
• Examined topic modeling algorithms (BerTopic, LDA etc.), implemented multimodal BerTopic model to cluster millions of posts (images and texts), optimized the topic representation, achieving a balanced performance on different metrics (coherence, topic overlap etc.).

“Drowsy Driving Detection under Different Circumstances with Convolutional Neural Networks (CNN)” with Pro. Colin Tan from NUS

• Developed the algorithm based on CNN to detect drowsiness with nearly 95% accuracy under different circumstances such as dark or bright.
• Built the backend classification server that performed the algorithm to handle the data and communicated data between detectors and actuators.
• Transferred data over Message Queueing Telemetry Transport (MQTT), RESTful APIs and stored in SQL databases.

• Designed an interview question generation algorithm, by combining knowledge graph and LLM, improving the depth and diversity of interview questions, smoothing the communication, and balancing the job requirements and the candidate’s experience.
• Built an automated resume parsing system, based on multiple LLM agents (i.e., GPT and Deepseek) and text processing tools (i.e., Tika and pdfplumber), achieving over 90% accuracy across diverse formats of resumes such as image-based PDFs and multilingual resumes.
• Demonstrated the effectiveness and potential socioeconomic biases of AI evaluation on candidates’ interview performance by using multiple methods such as hypothesis testing and regression to analyze the campus recruitment data.
• Examined millions of job posts and applications to illustrate and visualize the talent trends from multiple aspects such as talent flows across job types and cities and most popular skills.

• Independently investigated the whole supply chain of electrical vehicle, analyzing the competitive landscape and comparing various companies, and forming in-depth reports with more than 15,000 words.
• Built and managed a database for electrical vehicle to track the production and sales by country, firm, and model.