what is neuroflow fr?

NeuroFlow is a full-stack research prototype for brain-computer interface work. It ties together public EEG data, signal quality checks, classification models, a real-time inference service, and a browser dashboard into one reproducible pipeline. The README is explicit that version 3.0 is meant for research, teaching, and engineering validation only, not for medical diagnosis, treatment decisions, or high-risk device control. The data source is the PhysioNet EEG Motor Movement/Imagery dataset: five subjects, 64 channels, sampled at 160 Hz. Before modeling, the pipeline runs quality control that flags flat or abnormal channels, amplitude issues, and power-line noise, and checks label balance and power spectra around the C3, Cz, and C4 electrodes. Two model types are compared: a hand-built CSP plus LDA pipeline and a PyTorch EEGNet model, trained on two recording runs and tested on a third, held-out run. In the current small-sample setup, CSP plus LDA scored a mean accuracy of 60.00 percent against EEGNet's 50.67 percent, though the README cautions this result is specific to the current data and protocol and that generalizing across subjects still needs more subjects, repeated trials, and confidence intervals. A FastAPI backend serves the trained models over REST and WebSocket, streaming EEG packets in 100 millisecond chunks with a 4 second inference window. A browser-based dashboard, built with canvas rendering, shows live waveforms, per-subject quality reports, model comparisons, and system latency metrics, backed by a SQLite store of experiment records. The project can be run locally with a Python virtual environment and a PowerShell launch script, or built and run as a Docker container. It includes 25 unit and API tests and ships an OpenAPI spec for its endpoints. Data sources currently include PhysioNet replay and a mock hardware mode that simulates device packets, a Lab Streaming Layer hardware interface is planned but not yet connected. The project is released under the MIT License.