Prediction Intervals (PIs) are helpful to increase trustworthiness and user confidence in neural network regressions because they specify a range that contains the true estimate. Several methods for estimating PIs have been developed in the past, some of which rely on ensemble models whereas others rely on single neural networks. Although these methods have been successfully applied, there is still lots of room to improve the estimation accuracy of PIs. In this paper, we present a novel network architecture for estimating PIs with two interconnected networks. A Regressor network approximates the actual function; a Rater network estimates in tandem the PI boundaries around the ongoing regression by learning to estimate the deviation between the true output value and the computed regression. While the Regressor is trained with a training data set, there is no a-priori data set to train the Rater. To address this challenge, this paper presents a novel one-shot learning approach to train the Rater in parallel with the Regressor. Our evaluation on seven popular benchmarks shows that our proposed Regressor-Rater architecture constructs PIs that achieve high coverage probabilities of over 90% while being up to 55% narrower than the PIs generated by traditional approaches. Due to its small size and convenient design, our proposed architecture is particularly suitable to boost the trustworthiness of AIIoT applications.