New era of High Energy Physics experiments employing Streaming Readout techniques have to deal with very large amounts of data. This new approach is very challenging from the engineering point of view. Large detectors can easily eat up all the available bandwidth of the communication channel and key data could be lost. In this paper a Machine Learning approach for Data Reduction, an Autoencoder, is described and a low-cost hardware is built to implement it as a proof of concept. Different configurations of the autoencoder are implemented to study the tradeoff between model dimension, loss and inference time. In addition, a comparison with a standard lossless compression is made to highlight the benefit of the Artificial Intelligence supported algorithm.