research
Can LLMs Reason About Hate? An Evaluation of Hate Speech Detection and Classification in Reasoning and Non-Reasoning Models
Evaluated the ability of reasoning and non-reasoning LLMs to detect and classify hate speech, highlighting strengths, failures, and reasoning-driven improvements. Curated a custom dataset of over 100,000 hate speech and non-hate speech samples from across Youtube and X (formerly Twitter).
Evaluating Ethical Biases in Large Language Models
Examined bias behaviors in LLMs across socially sensitive scenarios and applied alignment methods such as RLHF and constitutional rules to improve fairness and value-consistent responses.
Phonic Forge: A Platform for Real-time Stuttering Detection and Personalized Therapy
Developed an LSTM-based speech disorder detection system trained on UCLASS and Sep-28k datasets, identifying five types of stuttering in real-time with 89.3% accuracy. Integrated audio preprocessing via spectral subtraction and Wiener filtering, with MFCC and spectral centroid feature extraction using Librosa. Fine-tuned LSTM and BERT models, selecting optimal architecture for deployment. Integrated Gemini 2.0 Flash for personalized speech therapy recommendations. Published in ICATES 2025.