In this work we report the results of a one year long performance evaluation of a solid state multi-sensor device calibrated with on field data, operating in a quantitative city air pollution scenario. Univariate enose calibration in controlled environment typically hold poor performance when tested in the harsh traffic operative conditions1. In a previous work, we proposed and tested the use of on-field data for the multivariate calibration of a multisensor device for estimating benzene concentration, with a conventional station providing the ground truth data2. Here, we present results obtained by this procedure when estimating NOx, NO2, CO concentrations over one year data acquisition campaign. Although showing a significantly worse relative estimation error with respect to benzene performances, results confirm that a neural calibration obtained with ten days long on field data segment can obtain optimal scores, with longer data segment being unable to obtain better performances.