Passionate about making an impact through AI.
Our goal is to identify the challenges and opportunities of building AI systems that would make a positive impact in the world.
Current Projects
AI and Security
This project investigates a new class of adversarial attacks on large language models—camouflaged jailbreaks—that use subtle, deceptive prompts to bypass safety filters.
Music and Young Minds
This project redefines explicit content detection in music by moving beyond the generic “explicit” label. We introduce a four-category framework—covering sexual references, violence, substance use, and offensive language—and develop a multi-stage pipeline for large-scale annotation. Our models significantly improve detection accuracy and reveal the shortcomings of current industry tagging practices, enabling more transparent and nuanced content moderation.
ARMOR
Augmented Retrieval for Mathematical Optimization & Reasoning.
This project enhances the mathematical reasoning abilities of large language models using Retrieval-Augmented Generation (RAG). By enabling models to dynamically access and incorporate relevant theorems, formulas, and definitions during problem-solving, we aim to improve their accuracy, explainability, and performance on complex, multi-step math tasks.
Code Switching
This project explores token-level prediction of code-switching in multilingual conversations—anticipating where speakers switch languages within a sentence. We develop both recurrent and transformer-based models using multilingual embeddings, achieving strong performance on benchmark datasets. The work advances code-switch prediction as a step toward more adaptive and inclusive NLP systems for multilingual users.
PLD-4: Detecting Paraphrased AI-Generated Text
This project introduces the Paraphrase-based LLM Detection Framework (PLD-4) to tackle the challenge of identifying machine-generated text, especially when paraphrased to evade detection.
By formalizing four detection tasks and evaluating models on benchmark datasets, we uncover the strengths and limitations of current approaches, providing a foundation for more reliable detection of layered and paraphrased AI-generated content.