A fast piecewise-defined neural network (PDNN) method is presented to produce accurate atmospheric temperature and humidity profiles from satellite hyperspectral infrared and microwave observations in all-sky conditions. The PDNN method relies on a novel classification approach and a principal component-based neural network function to better capture the nonlinear relationship between the spectral radiances and atmospheric state vectors. The algorithm was designed to only rely on satellite measurements. Large datasets were used for network training to make the retrieval robust to random errors in the reference data. For each retrieved profile, an effective quality indicator was obtained by training against the absolute value of the retrieval error. In addition, a reliable rain cloud flag was generated based on the scattering difference of cloud particles between the 50 and 118 GHz channels. Besides the independent reanalysis field data, the algorithm’s performance was also evaluated using radiosonde measurements. Preliminary validation shows that the best temperature retrieval occurred in the mid-troposphere (around 1.0 K). Depending on the reference data, errors in the boundary layer typically ranged from 1.8 K to 2.5 K. For water vapor, the retrieval error was less than 22% in the low-troposphere and less than 35% in the mid-and upper-troposphere when validated against the reanalysis field. The humidity error was approximately 10% lower when compared to radiosondes. The PDNN is used operationally to produce atmospheric temperature and moisture soundings from the Vertical Atmospheric Sounding System of the FengYun-3E satellite, an early-morning-orbit meteorological satellite launched in 2021.