Understanding what is relevant to a user is important for being able to deliver accurate and meaningful search results, engaging user experiences, and an efficient workflow. We present DeepBUFS, a prototype interactive system to support users make relevance decisions. DeepBUFS uses pre-trained deep learning models to classify relevance from user biometric samples. Preliminary results are encouraging, with deep learning models able to classify document topics from electrocardiograph (ECG) and electrodermal activity (EDA) with an average 71% accuracy. This suggests opportunities for new ways of interacting with users, including using biometrics to understand topical relevance and users' information and design needs. Further work is planned to expand these findings.