A childhood development system may capture audio and video in a home environment in which a child resides. The captured audio and the captured video may be filtered to remove audio and video corresponding to persons other than the child. Based upon information from the filtered video and the filtered audio, multiple scores can be calculated for the child using multiple trained machine-learning models. The scores can be compared with various stored activity score bundles. An indication of one or more compatible activities mapped to the activity score bundles that best match the child's scores is output.