The present work offers an architecture of a real-time intrusion detection system for IoT smart home gateway aiming at filling the research gap related to the contradiction between the high accuracy of deep-learning architectures and their application restrictions. The first task for this paper is to create a CNN-LSTM-Attention-based architecture ensuring not only high accuracy but also low-latency of operation in a streaming environment. The proposed methodology combines the use of convolutional layers for the local feature extraction, LSTM for sequence learning, and multi-head attention for the adaptive weight distribution in a four-thread pipeline processing the packet capturing, inference, alert generation, and flow management. Evaluation of the architecture is conducted on the NSL-KDD, CICIDS-2017 and TON_IoT datasets with the help of a set of preprocessing procedures, including the use of Min-Max normalization, RFECV feature selection and DeepSMOTE balancing for the class imbalance. The experimental results demonstrate the high accuracy of 99.61% on NSL-KDD, 99.21% on CICIDS-2017, and 98.84% on TONIoT. In the case of CICIDS-2017, the precision, recall, F1-score, and significantly decreased rate of false positives equal to 99.14%, 99.22%, 99.18%, and 0.43%, respectively, where the last one is 77% smaller than the baseline CNN, which makes the system extremely appropriate to be used in reducing alert fatigue in real-world applications. Also, the developed approach reaches an end-to-end latency of 0.37 ms per flow and an acceptable throughput of around 18,000 flows per second, meeting the criteria of real-time gateways under heavy traffic conditions. It has been shown via statistical significance tests that all obtained performance gains are statistically significant, where p<0.01. Overall, it can be stated that the study proves the importance of system-level optimizations and proper pipeline design over model complexity for building efficient, real-life, and high-performing IoT Intrusion Detection Systems.