In particle physics experiments, calorimeters are in charge of measuring the energy of incoming particles. In order to correctly estimate and evaluate the energy and other properties of these particles, a process, called reconstruction, is required. Because of the amount of collisions and the data-flow, reconstruction algorithms need to be time savvy. The nature of the problem seems appropriate for deep neural networks, yet the approach shows constraints. This paper presents the application to the calorimeter of LHCb in the first upgrade phase under the so-called Real Time Analysis framework, which is in charge of processing 30 MHz of data in real time, with its pros and cons.