A PhysNet neural network extracts your pulse waveform from raw webcam footage via remote photoplethysmography, no wearable required. Real-time HRV, heart rate, and respiration rate, from your face.
Rigorous empirical study of how unreliable LLM-as-judge is across GPT-4o, GPT-4o-mini, and Claude Sonnet. Key findings: judges flip their verdict in 15–22% of cases just from swapping A/B order; all three models prefer longer responses ~72% of the time (chance = 50%); human agreement tops out at k = 0.50 (moderate, at best). Bootstrapped CIs throughout. Because if your eval is broken, your model is too.
A full-stack platform built at the Augmented Intelligence Lab (University of Waterloo) to help newcomers navigate life in Canada. The backend is a Django REST API with JWT auth, PostgreSQL, and modular apps for organizations, events, mentorship matching, interpreter services, and volunteering. The mobile client is React Native (Expo) with Redux, react-native-maps, and i18next for in-app localization. Multilingual content is powered by Meta's SeamlessM4T model (via HuggingFace), supporting 30+ languages including Arabic, Farsi, Ukrainian, Tigrinya, Amharic, and more. Multi-contributor research project; related to the CHI '26 paper on capability-sensitive design for marginalized communities.