Deployment-Ready 5G NR CSI-Based Positioning using Machine Learning Baselines and FastAPI Microservice Implementation

Positioning is a key capability for intelligent wireless networks, enabling services such as geo-fencing, asset tracking, context-aware automation, and safety monitoring. Channel State Information (CSI), already estimated in 5G New Radio (NR) receivers for communication functions (e.g., equalization and link adaptation), can be repurposed as a sensing signal to infer user equipment (UE) location without additional positioning hardware. This paper presents an end-to-end prototype for 5G NR CSI-based positioning that (i) establishes reproducible classical machine-learning (ML) baselines for CSI-to-location inference and (ii) exposes the selected model through a lightweight REST microservice implemented with FastAPI, illustrating a practical pathway from offline training to online serving. We evaluate four baseline regressors (KNN, Random Forest, Gradient Boosting, MLP) for multi-output coordinate regression and corresponding classifiers for region-based positioning via grid discretization. Among tested baselines, RandomForestRegressor achieves the best validation performance with RMSE of approximately 2.4, while the tested MLP configuration fails to generalize in the high-dimensional flattened feature setting. The trained model and scaler are serialized and integrated into a FastAPI backend exposing prediction endpoints, demonstrating how CSI-driven inference can be operationalized as a positioning-ready service aligned with communication-and-sensing network analytics.