The tightening of global emissions standards necessitates innovative solutions for precise exhaust gas analysis. Conventional NOx sensors face challenges in gasoline engines, driving the need for alternative approaches. This study introduces an AI-based virtual sensor for NOx measurement, utilizing existing on-board diagnostics (OBD) signals to reduce hardware costs. A 4-cylinder gasoline engine was tested on an Engine-in-the-Loop (EiL) bench equipped with an exhaust gas analysis system for detailed NOx profiling. Using an XGBoost machine learning model, key parameters influencing NOx emissions post three-way catalyst (TWC) treatment were identified to develop a robust predictive model. Validation with data from a chassis dynamometer confirmed the sensor’s accuracy under real-world conditions. The virtual sensor enables accurate NOx measurements, validating or substituting physical sensors in case of failure, ensuring compliance with stricter regulations. This study provides a framework for training virtual sensors, addressing the challenge of modeling the catalytic aging effects on NOx conversion, detailing the model strengths and limitations while exploring future potential in emissions monitoring and control strategies.