With the rapid proliferation of IoT and AI-enabled smart devices across consumer and industrial domains, embedded IoT infrastructures centered around modem-enabled devices face rising security threats—particularly from internal actors. In such scenarios, conventional firewall-based solutions often fail to detect threats already present inside the network. Furthermore, most state-of-the-art machine learning models for network traffic classification derive conclusions only from offline experiments, lacking validation under real-time or embedded constraints. This article presents the Stacked Autoencoder-Integrated Network Traffic (SAINT) classifier, a layered learning framework that integrates autoencoders for feature compression and XGBoost for accurate anomaly classification. SAINT addresses key research gaps in dimensionality reduction, memory efficiency, and embedded deployability for IoT security. Designed for ESP32-S3 devices with SIM7670G modems, it applies unsupervised feature compression to reduce dependence on high-dimensional labeled data, achieving up to 50% input reduction for binary classification and 33.33% feature dimensionality reduction for multi-class classification. In deployment-oriented binary intrusion detection, SAINT achieves an attack-class recall of 95.08% with a false positive rate of 6.97%, outperforming existing approaches evaluated under the same experimental conditions and demonstrating effective real-time anomaly detection on resource-constrained IoT platforms.