Modern societies are rapidly entering an era of population aging, accompanied by a growing number of elderly individuals living alone. This demographic shift presents serious challenges in eldercare, as older adults are vulnerable to life-threatening emergencies such as cardiac arrest, stroke, or falls that may remain unnoticed for extended periods, as well as early signs of cognitive decline that are often detected only after irreversible progression. Existing monitoring approaches, including wearable devices, environmental sensors, and cloud-based systems, typically rely on single-modality inputs and suffer from limited accuracy, scalability, and long-term practicality. To address these limitations, we propose a Transformer-based multimodal on-device AI system that integrates three complementary sensing tasks: vibration-based life-presence detection, multi-sensor behavioral anomaly detection, and speech-based cognitive screening. A Transformer-based fusion architecture is employed to model cross-modal relationships and enable reliable real-time inference within a single, privacy-preserving device. In a pilot study, the proposed system achieved a sensitivity of 94.6% for life-presence detection, an AUROC of 0.92 for behavioral anomaly detection, and an AUC of 0.91 for cognitive-risk prediction. The full multimodal fusion further improved performance, reaching an AUC of 0.95 with an average of 0.8 false alarms per day, a 5.3% miss rate for high-risk events, and an average detection time of 19.6 min, significantly outperforming single-modality baselines. These results demonstrate the technical feasibility and societal value of a practical, on-device AI solution for enhancing safety and cognitive health monitoring among vulnerable elderly populations in aging societies.