Research Project: Neural Network for Underlying Law Understanding and Generalizable Reasoning.Supervisor: Davi Geiger, Associate Professor, Courant Institute of Mathematical Sciences, New York UniversityRonen Basri, Researcher, Meta Platforms Inc. | Full Professor, Weizmann Institute of ScienceOur team improves the neural network’s generalizability, underlying law understanding, and reasoning in predicting accurate output for input not covered by training data, seeking alternatives to the approximation-based approaches used in traditional neural networks with limited accountability. I developed a novel neural network architecture featuring multiple activation functions from a library of component functions and phase space mapping via discrete Fourier transform, capable of learning the exact hidden function between physics variables. Implemented the neural network architecture and optimizer with TensorFlow. Our model is capable of predicting accurate output in an indefinite domain of multi-layer composite functions.

Research Project: YIZHE: An AI-Facilitated Post-clinic Service Chatbot.Supervisor: Yuntao Wang, Associate Professor, Department of Computer Science and Technology, Tsinghua UniversityOur team investigates the effectiveness of different techniques (prompt engineering, RAG) in integrating Large Language Model (LLM)-based chatbots for post-clinic services by conducting user experiments. I integrated OpenAI API to develop the GPT-4o-based customer service chatbot by designing the instructions, response template, and few-shot training data generator with state-of-the-art methodologies in prompt engineering for medical agents, including few-shots learning, discrete prompting, and prompt generation. Co-developed the human-AI collaboration system involving human doctors, patients, and AI agents interacting through instant messaging platforms, with AI agents for instant basic care and human doctors for holistic diagnosis. Co-designed and conducted user experiments with Enjoy Dental Clinic to study patients’ perception of chatbots by deploying chatbots with instant message platform API and designing questionnaires. Co-authored a 19-page paper. Our results validated that RAG chatbots outperforms human doctors in user satisfaction by 3%.

Our group employs mathematical models and algorithms to make investment decisions of equities. I visualized the performance macroeconomic indicators for stock-index futures investment using Python (Numpy, Matplotlib...) to provide a straightforward illustration for decision-making of fund managers. Implemented a long-short term memory (LSTM) neural network stock price predictor in TensorFlow. Determined the dominant fundamental analysis factors for selecting Microcap Index components using Wind market data terminal, and reached an counter-intuitive conclusion that a single factor outperforms combinations. Created "n-day max and min" indicator that determines the longest interval the stock price hits extreme values, and backtested its performance on CSI 300 Index to understand the impact of historical extremes on future prices. Developed a stock data crawler using OpenCV and PyAutoGUI that saves stock data locally for future use.