We present Trivana, an on-device mental health mobile application that enables holistic, proactive, and privacy-preserving support. The system integrates multimodal user inputs, a forecasting model for anticipating mental health risks, and a local LLM-based chatbot for personalized interventions. By unifying sensing, forecasting, and interaction within a single pipeline, Trivana overcomes fragmentation in existing apps while ensuring privacy guarantees through fully on-device execution.