The brain computer interface (BCI) is a nonstimulatory direct and occasionally bidirectional communication link between the brain and a computer or an external device. Classically, EEG-based BCI algorithms have relied on models such as support vector machines and linear discriminant analysis or multiclass common spatial patterns. During the last decade, however, more sophisticated machine learning architectures, such as convolutional neural networks, recurrent neural networks, long short-term memory networks and gated recurrent unit networks, have been extensively used to enhance discriminability in multiclass BCI tasks. Additionally, preprocessing and denoising of EEG signals has always been key in the successful decoding of brain activity, and the determination of an optimal and standardized EEG preprocessing activity is an active area of research. In this paper, we present a deep neural network architecture specifically engineered to a) provide state-of-the-art performance in multiclass motor imagery classification and b) remain robust to preprocessing to enable real-time processing of raw data as it streams from EEG and BCI equipment. It is based on the intertwined use of time-distributed fully connected (tdFC) and space-distributed 1D temporal convolutional layers (sdConv) and explicitly addresses the possibility that interaction of spatial and temporal features of the EEG signal occurs at all levels of complexity. Numerical experiments demonstrate that our architecture provides superior performance compared baselines based on a combination of 3D convolutions and recurrent neural networks in a six-class motor imagery network, with a subjectwise accuracy that reaches 99%. Importantly, these results remain unchanged when minimal or extensive preprocessing is applied, possibly paving the way for a more transversal and real-time use of deep learning architectures in EEG classification.