Intrusion detection in IoT networks poses significant challenges due to the limited computational resources of embedded devices and the growing sophistication of cyberattacks. In this work, we propose and implement a lightweight Deep Neural Network (DNN)-based intrusion detection model directly on an embedded router platform. The model is trained on the IoT-23 dataset and deployed on the MediaTek AI7688 system-on-chip running OpenWRT. Two experimental scenarios are evaluated: offline testing with pre-collected data and real-time packet capture using libpcap. Results show that the model achieves a high detection accuracy of 99.9%, while requiring only 8.6 KFLOPS per inference, 0.71% average CPU usage, and 6.64 MB RAM. Furthermore, the model delivers predictions within 1.42 ms, ensuring real-time responsiveness. These findings demonstrate that embedded routers can effectively integrate AI-based intrusion detection without compromising networkingperformance. The proposed approach highlights the practicality and scalability of deploying lightweight DNN models on open-source router platforms, paving the way for secure and intelligent IoT infrastructures.