This talk presents a recently patented sensor for measuring turbidity in water. The sensor works by photographing laser light scattering in the water and processes the images using a multi-input, multi-output neural network running on an efficient low-power processor (Greenwaves Technologies GAP). Thanks to the low energy consumption, the sensor can be put in a stormwater well for several years while communicating through NB-IoT. The neural network’s regression head processes the captured images by combining the distance to the surface with the photographed image, resulting in two outputs: the predicted turbidity and an error indicator to determine the similarity of the photographed image to the data it has been trained on. A significant challenge faced during the development of the sensor was obtaining training data with accurate ground truth information and conditions that closely resembled real-world scenarios, as data collected in stormwater wells is often not well-distributed and therefore also not suitable for training. Additionally, quantizing the neural network model to reduce memory usage while maintaining accuracy is a challenge.