Diet tracking via self-reports or manual taking of meal photos might be difficult, time-consuming, and discouraging, especially for children, which limits the potential of long-term dietary assessment. We present the design and development of a proof of concept of an automated and unobtrusive system for diet tracking integrating: a) a social robot programmed to automatically capture photos of food and motivate children, b) a deep learning model based on Google Inception V3, applied for the use case of image-based fruit recognition, c) a RESTful microservice architecture deployed to deliver the model outcomes to a platform aiming at childhood obesity prevention. We illustrate the feasibility and virtue of this approach, towards the development of the next-generation computer-assisted systems for automated diet tracking.