At Samsung PRISM R&D, I fine-tuned RoBERTa-Large for AI-generated text detection, reaching 98.0% evaluation accuracy on the HC3 dataset. I then used knowledge distillation to reduce the model from 355M to 125M parameters; the RoBERTa-Base student achieved 93.5% accuracy and 3.18× faster CPU inference.
I also built Symptom Analyzer ML, a Streamlit app that pairs a Random Forest disease classifier with locally hosted Llama 3 explanations grounded in the predicted class and symptoms. My projects include a Java and Spring Boot REST API with JWT authentication, role-based access control, and BCrypt password hashing.

