Failure of the urban water supply is first detected by the users, but municipal complaint systems tend to store complaints as individual text files and attachments. In this paper, the design of AquaFix, a civic platform for gathering text, images, and GPS data, followed by a review from the municipalities based on the roles, is provided. It's an applied system architecture, not a new neural-network architecture, and the use of the task-specific deep-learning artifacts is evidenced. The documented stack comprises React, Node.js, Express.js, MongoDB, JWT authentication and role-based access control. 
The model audit consists of five categories: water-complaint classification by BERT, pipe-leakage segmentation by YOLOv8, spilled water detection by YOLOv8, floating debris detection by YOLOv8 and waterlogging classification by ResNet-50. The original-test accuracy reported by the BERT is 100.0\%, and the challenge accuracy is recalculated independently by BERT as 99.82\% (565/566). ResNet-50 has an independently calculated test accuracy of 91.9847\% (241/262). The pipe-leakage and spilled-water measurements are epoch-50 validation values; the floating-debris measurements were not recalculated separately, but are recorded in the artifact test CSV. These model artifacts are clearly distinguished from live Express.js inference, multimodal fusion, learned urgency scoring and end-to-end deployment, which are not confirmed or further research.

complaint classification,deep learning,water infrastructure,smart cities,computer vision