Currently, existing voice-assisted smart homes are based on cloud-based computation, which gives rise to several concerns regarding data privacy, latency, and dependency on the network. This paper proposes a novel, secure, and efficient smart home system, namely, Localized User-initiated Machine Intelligence (LUMI), which ensures data privacy, ultra-low-latency voice interaction, and real-time device automation through local speech processing, semantic reasoning via Large Language Model (LLM), and IoT device management through a completely local architecture. Unlike existing cloud-based smart homes, in the proposed system, all computation occurs locally, ensuring data privacy, ultra-low-latency voice interaction, and real-time device automation. By employing local computation of the Mistral LLM, the proposed system can efficiently translate voice commands into device management messages through the MQTT protocol. Experimental results validate the proposed system as a novel, secure, and efficient solution, ensuring 96.7 % accuracy and 412 ms latency.