Selected Works

An explainable Twitter list recommendation system built on a knowledge graph with 5 relation types (user–user, user–topic, user–list, list–topic, topic–topic). Combines an A* baseline with Graph Neural Network embeddings achieving 0.27 Precision@10 and 0.58 Recall@10 — a significant improvement over the baseline.

The first comprehensive security vulnerability scanner for Model Context Protocol servers. Detects prompt injection, tool poisoning, rug pull attacks, server impersonation, and 10+ other critical risks using a multi-layered approach combining keyword detection, semantic analysis, and LLM evaluation.

Predicts the next music genre a user might enjoy based on their mental health diagnosis (depression, PTSD, anxiety, bipolar) and Twitter-derived listening history. Uses LSTM and GRU sequential models with t-SNE visualization to explore connections between music preferences and mental health.

A comparative study of 7 regression algorithms (Random Forest, Gradient Boosting, KNN, SVR, Linear, Ridge, Lasso) for predicting house sale prices. Includes full data preprocessing pipeline and feature selection. Gradient Boosting achieved the highest accuracy.

A constraint satisfaction problem solver for the Kakuro puzzle implemented in Java. Uses CSP backtracking with AC3, MRV selection, LCV ordering, and forward checking to efficiently solve the puzzle while satisfying all row/column sum constraints.

A compiler for Jython — a hybrid language combining Java and Python syntax. Implements lexical and syntactic parsing, common error detection, a listener pattern, and an inheritance tree walker using ANTLR grammar definitions.

A computational framework for personalized music recommendation based on mental health states (anxiety, depression, PTSD). Uses mixed-effect models and propensity score analysis to study the causal impact of music listening on stress, published in ACM Transactions on Information Systems.

A CNN-based seizure detection model trained on EEG signals from the PhysioNet CHB-MIT dataset. Extracts 15 signal features per channel, applies 4 data augmentation techniques to handle class imbalance, and achieves 87% accuracy with 0.94 recall using 3 Conv1D + 2 Dense layers.