Soft sensors are models that allow estimating the values of a variable based on other process information, without having to measure this variable directly. The main benefits of soft sensors are (1) they represent a low-cost alternative when compared to physical sensors, (2) they can work together with physical sensors, including to identify when they fail, (3) they allow implementation on existing devices, and (4) they provide real-time estimates, being an option for measurements where physical sensors depend on time-consuming analysis. In this tutorial, we are going to learn how to develop a data-driven soft sensor using Python taking into account data-driven techniques such as neural networks, decision trees, and other regression techniques. Besides the soft sensor development, we will discuss 10 good practices to develop safer, more reliable and more efficient models. The good practices involves the understanding of the risk of not taking care of models extrapolation, the effect of data quality and data quantity while building the models, and the importance of error distribution besides general metrics. After building the models, we will discuss good practices to improve efficiency, how to monitor performance and how to perform models calibration. Real data from industrial processes will be used in this hands-on tutorial.