Dementia care frequently requires continuous supervision, a demand that becomes increasingly difficult when caregivers must balance additional responsibilities such as work or school. This challenge became personal when my grandfather, diagnosed with mild dementia, wandered from home while alone, revealing a critical gap in existing care technologies. While existing tools aim to support dementia patients, they often rely on manual patient input, an expectation that is unreliable for individuals with impaired memory and judgment. To address this limitation, I developed an AI-based voice-interactive caretaker to assist patients through natural, everyday conversation rather than commands or interfaces. Operating through a physical form factor such as a wearable or tabletop device, the system continuously listens and interprets patient speech, determining whether it reflects daily activity, medication intake, an emergency, or casual conversation. When relevant, activities are inferred and logged, medication intake is recorded while reminders are seamlessly integrated into ongoing dialogue, and emergencies are detected, assigned a severity score, and communicated to caregivers via real-time SMS alerts; otherwise, the system continues to facilitate conversation unobtrusively. To evaluate system viability, twelve scripted scenarios representing realistic and edge-case interactions are executed across ten trials each, measuring speech recognition latency, inference and response time, data logging efficiency, emergency detection accuracy, and severity scoring efficiency. Quantitative analysis of these metrics assesses reliability and highlights opportunities for optimization. As dementia prevalence continues to rise, this research explores how human-centred AI systems can enhance patient safety, reduce caregiver stress, and provide an ethical, scalable framework for home-based care.