ABSTRACT Ground-shaking maps provide a spatial representation of the impact of a seismic event in terms of ground-motion parameters (GMPs), especially useful in the context of seismic monitoring and civil protection operations. Algorithms used to compute these maps usually rely on seismic source parameters to steer the interpolation process and consequently are limited to operate in near-real time. The present work introduces a novel algorithm that combines neural networks with the multivariate normal distribution method to reconstruct ground-shaking maps using only data available in real time, improving on previously proposed algorithms. The core idea of the proposed algorithm is to maintain the structure proposed by ShakeMap while removing the dependence on the source parameters, imposed by the use of ground-motion prediction equations, by replacing them with an appropriate neural network working on the GMPs recorded in real time at the seismic stations. The overall workflow of the method and the details of the neural network architecture and training are described. A model trained on synthetic and recorded data to target seismic events affecting the Italian territory is tested using the 2016 Norcia, Italy, earthquake showing the method reconstruction capabilities, its robustness to noise and to network geometry changes, and its real-time potential.