neurOS-v1

Neural data work is fragmented: one tool acquires from a device, another filters, another trains a model, and nothing shares an architecture. neurOS-v1 makes the whole path one coherent, extensible platform — from hot-swappable device drivers and real-time streaming to a multimodal foundation model and mechanistic-interpretability tooling — organized as a 10-package monorepo so every concern stays independently usable.

• 10-package monorepo: core, drivers, models, foundation, neurofm, mechint, cloud, sourceweigher, ui
• Unified hot-swappable driver API across 16+ modalities (EEG, ECoG, EMG, ECG, EOG, fNIRS, GSR, calcium imaging, video, audio, motion, respiration, phone)
• Async agent orchestrator coordinating device, processing, model, and fusion agents in real time
• Pluggable processing pipeline: filters, feature extraction, adaptation, and health monitoring loaded at run time
• Registry-driven model zoo: EEGNet, CNN, LSTM, Transformer, DINOv3, attention-fusion, plus classical SVM/KNN/RF/GBDT
• neuroFMx multimodal foundation model with per-modality tokenizers and a Mamba backbone
• Constellation demo: multi-modal sync to NWB/Zarr, WebDataset export, and distributed training

• Motor Imagery Classification
• Multimodal Fusion
• Foundation Models Demo
• DINOv3 Neuroscience Experiments (Allen atlas, CREMI, SNEMI3D, cell tracking)