Project

This project investigates the use of Generative AI (GenAI) for Cognitive Task Analysis (CTA), focusing on task decomposition, decision modeling, risk assessment, and automated report generation. By leveraging LangGraph and Google Gemini AI, the project aims to enhance structured reasoning and assist professionals in complex cognitive workflows.

Through qualitative evaluation and bias minimization techniques, the research ensures that AI-generated reports are accurate, explainable, and applicable across multiple domains.

Phishing attacks are a major cybersecurity threat, and blocking spam emails is not enough—we need to understand how scammers persuade people to fall for their tricks. This project aims to label phishing emails based on tactics like urgency, authority, and importance, but since no ready-made labels exist, I used zero-shot learning with Llama 3 and the ChatGPT API to generate them. To keep results stable, I set the temperature to 0 and then used K-Means clustering to group similar labels, identifying eight key persuasion strategies. This method successfully classifies phishing emails, helping improve detection systems and cybersecurity defenses. This AI-driven approach can change how we detect phishing attacks, making online security smarter and more proactive.

The project utilizes a small phishing dataset to generate new phishing emails by using seq-to-seq (GRU, LSTM) models, combining transfer learning using the GloVe matrix. Beam search was used to decode the model to generate new text.  Then, the generated text is evaluated by the spam classifier created by machine learning (XG-Boosting, SVM, Random Forest) The model is able to use impersonation strategies generating messages. Click here to see more details.

The objective of the research is to automate the diagnosis process of the engine failure problem. The research use truck warranty claims data to create classifiers by using machine learning and deep learning. To resolve the problem with 2 million text entry with 200+ classification labels, hierarchical structure classification has been applied to solve multiple levels of claim labels to achieve high performance. Network analysis was applied to identify complicated relationships between labels. The research outperforms the existing tools and provides insightful suggestions for further exploration in data collection.

In this project, I implemented a classic agent-based simulation problem, El Farol restaurant, under the mesa package and substituted the original agent with a reinforcement-based model (Instance-based learning). Further, I replicated a paper and experimented with the effect of homogenous and heterogeneous groups on the overall bar attendance dynamics. The code was also submitted to contribute to the mesa simulation package as an example skeleton. Please see the GitHub page for more details. 

The project explores two optimization methods(GA & IPA) in the portfolio optimization problem to compare with pure random search. The question addressed in the project is to figure out the best combination of a group of stocks with balance between volatility(risk) and expectation(reward). Five stocks are chosen to validate the algorithms. GA and IPA are able to select the best investment strategies with different risk preferences and narrow down the available choices from pure random search. Click here to see more details.

In the project, we design a speech-only text editor to address some of issues that were found in existing speech-based editing systems. The focus was on improving navigation and reducing the complexity of existing interfaces. A simple editing task was designed in order to identify the limitations of Google’s exiting Voice Typing editor and to compare the performance to our design. Click here to see more details.