One of the major problems in advancing Internet of Things (IoT) technology is the need for fast complex data processing, feature extraction and classification tasks.^[1]  A further challenge is the need for powerful computing facilities as close as possible to the physical interface, implemented in edge computing systems, to decrease the power management, scalability and sustainability of cloud computing infrastructure.^[2] A notable example of edge  system is Brain-Computer Interface (BCI), that is a rapidly emerging field with applications in domains as prosthetic devices, robotics, communication technology, and security.^[3] To efficiently interface the brain with electronic devices for the recording, and possibly the in loco processing of signals, a major problem is represented by the real-time processing of raw neuronal signals, which imposes excessive requirements on bandwidth, energy, and computation capacity, often asking for a severe pre-processing task.[4] Neural networks can be employed in edge computing solutions for classification tasks.^[5] However, to leverage the inference capabilities of these learning machines, time-series data must first be flattened and then encoded into spike trains. This step is not trivial, as it introduces an additional layer of complexity in the processing chain and significantly reduces the temporal resolution originally present in the raw time-series data.^[6] Among various strategies developed to overcome these issues, cluster-assembled thin films are here proposed as novel hardware data processing solutions to efficiently perform reprogrammable computation and signal processing on the edge of the physical system under investigation. Metallic cluster-assembled materials, deposited by Supersonic Cluster Beam Deposition (SCBD), are characterized by a complex network composed by a high density of defects and grain-boundaries.^[7] These nonlinear electrical properties^[7–9] can be exploited for the development of novel paradigm of computation, as reconfigurable nonlinear Threshold Logic Gates.^[10] The engineering of the metal cluster-assembled thin films can be further developed and implemented in hybrid computing architectures, used for processing signals recorded on edge.^[11] As a case of interest, we report an in materia approach to perform on edge real-time series classification tasks,^[12] based on cluster-assembled thin films and a time-series analysis method proposed by Fulcher et al.^[13]. We used a nanocomposite resistive switching device, based on gold and zirconia (Au/ZrO_x) thin film,^[14,15] to project the input time-series into a higher dimensional-space, allowing the resulting output time-series to be further analyzed by a linear classifier. We demonstrated the potential of this method to classify with high accuracy and in real-time neuronal traces, recorded by a neural probe in the barrel cortex of a rat, in spontaneous and elicited conditions. The classification was carried out with limited datasets for training and memory storage, and characterized by higher interpretability and accuracy with respect to artificial neural networks used on the same neuronal traces^[16]. The proposed methodology is well-suited for its extension to other neuromorphic devices, in particular to all those systems with a fast response to stimuli to classify highly resolved temporal time-series.