Internet-of-Things (IoT) helps create smart systems by allowing physical environments to be controlled or monitored by users through extensive sensor networks. Such networks are heavily resource constrained due to the low processing power and decentralized placement of sensors. Having been implemented in several backbone sectors such as industry, healthcare, etc, they often deal with sensitive information. Thus, there is an imminent need to secure IoT networks using lightweight intrusion detection systems. To this end, existing works that use BiLSTM for network intrusion detection are explored. Experiments involving the UNSW-NB15 dataset are recreated, and the results are compared. Early stopping is implemented to improve model performance and enable better analysis of the results. Finally, a novel machine learning and feature engineering-based intelligent network intrusion detection framework, dubbed FEIIDS, is proposed. Its working is explained in depth and its time complexity is analyzed. It is tested on the UNSW-NB15 dataset and found to be competitive with existing works, achieving classification accuracies up to 98.8%. Discussions on the results reveal that FEIIDS is more stable, transparent, and lightweight than deep learning models. In fact, FEIIDS can process 2,032,038 records of the dataset in 280.44 seconds, whereas BiLSTM requires 1901.2 seconds to process just 14,000 records. A case study on FEIIDS’s real-world application is also presented. Finally, it is concluded that FEIIDS is highly suitable for use as a real-time intrusion detection system in resource-constrained IoT networks, and its possible future applications are listed.