Mental health concerns like stress and anxiety are growing rapidly on digital platforms. Traditional diagnostics are often reactive and unscalable. Existing automated systems typically focus on single modalities, limiting detection accuracy. This project proposes a multi- modal framework for early mental health risk assessment. The system integrates typing behavior analytics (speed, latency, errors) with natural language analysis using BERT-based transformers. These streams are fused via ensemble learning to comprehensively understand emotional states. The model generates real-time risk scores to provide early warn- ings. Designed for real-time deployment, the framework includes data preprocessing, fine-tuned training, and fusion-based classification. Fu- ture enhancements include physiological signals and privacy-preserving methods. Ultimately, MindWatch establishes a foundation for intelligent, scalable, and proactive mental health monitoring.