Researched partial credit estimation methods for student-written database queries in assignments and exams, to reduce TA grading workload by over 50% and improve grading outcomes for the 65% of questions that previously received a score of zero. Integrated AI agents into the grading pipeline using Retrieval-Augmented Generation (RAG), Fine-Tuning, and prompt engineering, achieving an MAE of 6% and an RMSE of 12%. This thesis project is advised by Professor Walid G. Aref and Professor Hisham Benotman at Purdue University.

• Python
• Pandas
• Numpy
• SciPy
• Scikit-learn
• SQL
• Matplotlib
• PyTorch
• LangChain

View Project

The strategy first uses PCA to extract market risk factors from the correlation matrix of day-to-day equity returns, keeping the top eigenvectors to represent systematic movements. Thus, each stock’s return is decomposed into systematic factors and a residual (idiosyncratic) component. These residuals are modeled as Ornstein–Uhlenbeck mean-reverting processes, and trading signals are generated when a stock’s residual deviates significantly from its estimated equilibrium, measured through z-score.

• Python
• Pandas
• Numpy

View Project

This project is a Vision-Language Model (VLM)–enhanced reinforcement learning framework for autonomous driving, leveraging pretrained models like CLIP to provide expert-like guidance to RL agents. By integrating VLM-derived feedback—such as action suggestions, safety scores, and scene understanding—the system improves sample efficiency and reduces reliance on pure self-exploration. The method introduces three key components: Value-Margin Regularization, Advantage-Weighted Action Guidance, and VLM-based reward shaping, all designed to steer RL agents toward safer and more effective driving behaviors. Experiments in the CARLA simulator show that VLM-guided agents achieve higher route completion, better safety, and lower energy consumption compared to standard RL approaches.

• Python
• Pandas
• Numpy
• HuggingFace

View Project

This project investigates whether compressed large language models (LLMs), such as distilled or quantized versions, can effectively detect fake news while reducing computational costs and time to make feasible even on resource-constrained devices. Using the LIAR dataset, we compared compressed models against full-sized counterparts on accuracy, efficiency, and explainability, applying techniques like prompt engineering and retrieval-augmented generation (RAG).

• Python
• Pandas
• Numpy
• PyTorch
• HuggingFace

View Project

This project investigates the application of Deep Reinforcement Learning (DDPG) for optimizing routing decisions in SDNs on grid, random, Internet-MCI, and fat-tree topologies.

• Python
• Pandas
• Numpy
• NetworkX
• Matplotlib
• PyTorch

View Project

Extracted different metrics and then implemented PCA and correlation analysis on different time periods. Used useful and non-correlated data to run classifiers such as Logistic Regression, kNN, Naive-Bayes, etc to predict different race outcomes such as Podium Winners, Point Scorers, etc.

• Python
• Pandas
• Numpy
• Scikit-learn
• SciPy
• PyTorch
• Matplotlib

View Project

During the pandemic, we developed a project called Classmate Plus that lets teachers and students make online classes more interactive and engaging.

• JavaScript
• HTML/CSS
• Python
• Django
• Docker
• Flask
• SQL

View Project