Accurate, high-resolution air quality data are crucial for understanding environmental health risks; however, the cost and complexity of maintaining dense, reference-grade monitoring networks remain a significant barrier. This study presents the first city-wide evaluation of next-generation air quality sensors in Zagreb, Croatia, involving 35 sensor locations, one local reference-grade station, and three national reference stations that measure PM₁₀ and NO₂. Sensor performance was evaluated against reference data under various meteorological and temporal conditions. To better understand sensor drift and measurement bias, we developed machine learning (ML) calibration models (XGBoost) using spatiotemporal features, ERA5 meteorological variables, and traffic proxy indicators. The models significantly improved accuracy, reducing the root mean squared error (RMSE) by up to 82%, with the greatest improvements observed during pollution peaks. A rolling Root Mean Square Error (RMSE) approach was introduced to track model degradation over time, revealing that recalibration was typically needed within 1–6 months. Our findings demonstrate that, with proper calibration and maintenance, sensor networks can serve as reliable and scalable tools for urban air quality monitoring, capable of supporting both public health assessments and informed decision-making.