- Researching and developing a retrieval-augmented multimodal large language model (LLM) for understanding and question-answering on multimodal financial data through information retrieval, facilitating decision-making in business and investment- Final Year Project

• Researching compression of large-scale models with AutoML techniques under Prof. Bryan Low's MapleCG lab• Designing algorithms for and experimenting on task-specific iterative pruning of LLM, reducing model size to 30% while maintaining performance• Building a novel adversarial knowledge distillation strategy, achieving comparable result as SOTA methods• Implemented Bayesian Optimization on input tuning, improved generation of teaching examples for student model

• Built and trained computer vision models (VGG16, EffcientNet, etc.) for brain tumor diagnosis, achieved 98% accuracy• Developed transformer-based solutions for cancer predictive modeling and drug discovery with electronic health records (EHR)• Fine-tuned a multimodal LLM from LLaVA with DeepSpeed ZeRO to build a chatbot for healthcare• Integrated RAG with the LLM, reduced hallucination• Engineered the backend to host the LLM on a distributed cloud computing service, enhancing scalability and reliability

• Engineered a real-time monitoring app for MRT using C# and WPF, streamlining the visualization of hardware states and logs• Established and maintained robust and resilient connections with multiple servers via Modbus TCP/IP and multithreading• Implemented decoding functions to transform raw server data into intuitive notifications according to custom configurations• Uniformly integrated data models with frontend XAML UI components using MVVM pattern. Revamped and refactored the codebase from previous interns to fit MVVM and improve data consistency. Ensured code quality, efficiency, and extensibility• Coded and dockerized RESTful APIs in .NET to seamlessly integrate frontend and backend, enabling dynamic data exchange• Collaborated with collogues and conducted code reviews via Git and GitHub team development tools e.g. PRs, CI/CD, etc.

Organized events like China Submit to help bridge students with business leaders.

• Lectured a 2-hour tutorial weekly to recap the concepts and properties of different DS and Algorithms• Guided students to solve problems with efficient algorithm in-class to train their way of thinking as a programmer• Performed Q&A sessions for students in a timely manner and provided detailed feedback on their assignments

• Lectured a 2-hour tutorial weekly to recap the concepts of the programming foundation module• Guided students to solve difficult programming problems in-class to train their way of thinking as a programmer

• Built machine learning model with PyTorch to discover students’ proficiency level on knowledge concepts and predict their future exams scores, based on the Neural Cognitive Diagnosis framework in a published paper• Adapted the machine learning model to fit continuous real-world data instead of only discrete pseudo data• Integrated functionalities to detect input file format, mapped training set to the standardized format, and saved corresponding information for validation and testing• Tuned the machine learning model parameters by conducting convergence analysis given batch size, learning rate, and data normalization condition• Improved the model by adding time-series data identity and adjusting the weight on the training set based on data collected time and question type. Achieved accuracy of 80% finally.