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).
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.