This work advances the machine learning based GNSS positioning by demonstrating the potential of neural network to adapt to different challenging regions, such as dense urban and urban canyon. The adaptive nature of neural networks make it a promising alternative to solve the severe multipath errors in different urban scenarios. Combining both the region adapted neural network and the traditional Kalman Filter based algorithm, better measurement and positioning are performed for a specific urban area. This work integrates the region adapted neural networks with MediaTek's NeuroPilot framework to enable efficient computing on mobile devices. Our implementation deploys a 5-layer Multilayer-Perceptron (46K neurons) on MediaTek's APU 2.0 and achieves real-time performance for typical urban cases of 15 available satellites (4.5ms and 2.7MB memory access). The actual measurement results conclude that neural networks can adapt to urban areas and thus achieve better positioning accuracy (24% and 28% positioning accuracy improvement for Shenzhen and Hsinchu city respectively). Challenging urban areas, including canyons, buildings and dense foliage make the GNSS positioning an extremely complicated task. One of the major interference of GNSS signals is the multipath error, where signals are reflected, refracted and/or absorbed. Such phenomenon results in multiple paths of arrival for a GNSS receiver. Moreover, these harsh conditions get even worse since the environment complexity are highly varied across different cities, such as San Francisco, ShenZhen and HsinChu. This work demonstrates the capability of neural network to adapt to different regions. The adaptive nature of neural networks make it a promising alternative to solve the severe multipath errors in different urban scenarios. In the most of traditional designs, the Kalman Filter has been widely applied to GNSS positioning receiver. The filter performs smoothing, estimating and predicting in a sense of least-square methodology. The objective is to minimize the sum of squared residuals (the difference between a measured value and the fitted value provided by a model). With the assumption of Linear Time-Invariant (LTI) system with Gaussian noises, the result obtaining from Kalman Filter can be considered as an optimized one. However, when performing the positioning for multiple urban scenarios. The major challenge is to continuously adjust the Kalman Filter parameters to adapt to the environment of operation. To improve the accuracy of positioning, this work integrates the region adapted neural networks to refine portions of the GNSS receiver data. These enhanced GNSS receiver data are later feed into Kalman Filter. We propose a new method by applying a 5-layer Multilayer-Perceptron (46K neurons) to classify the quality of satellite signals. In our implementation, such signal quality classification is used to weight the GNSS receiver data. Based on the weighted GNSS receiver data, refined measurements are obtained and further served as the input to the Kalman Filter. As a real implementation of the system, the complexity of the neural network is kept small enough to be fitted into a mobile device which has memory and computation constraint. The proposed region adapted neural networks is evaluated in MediaTek's mobile platform with its NeuroPilot framework. In our implementation, a 5-layer Multilayer-Perceptron (46K neurons) is deployed on MediaTek's APU 2.0. With the processing cores of APU 2.0, the major neural network workloads can be offloaded from the traditional positioning unit. The resultant performance reaches 4.5ms and 2.7MB memory access (for typical urban cases of 15 available satellites), which fulfills the real-time requirement for a positioning system. Moreover, the real evaluation shows 24% and 28% positioning accuracy improvement compared to standard Kalman Filter in the urban areas of Shenzhen and HsinChu city, respectively. In conclusion, this work explores the benefits of using region adapted neural network, and integrate it with Kalman Filter based method to mitigate the effect caused by the corrupted measurements (e.g., multipath errors) in challenging urban areas.