Principles
Local-first
Simple local install:
pip install oneopen-ml-studio
oneopen start
Local mode requires:
No external cloud account
No mandatory registration
No external data upload
No extra infrastructure beyond Python
SQLite + local filesystem
Self-hosted by default
Deploy on a workstation, internal server, VM, Docker Compose, Kubernetes, or private cloud.
Framework-neutral
Avoid hard dependency on a single ML stack. Shipped adapters: Ultralytics, LibreYOLO, Torchvision, scikit-learn, Hugging Face prepare, modality packaging. Direction (not all implemented): TensorFlow/Keras, XGBoost, LightGBM, spaCy, Detectron2, MMDetection, Whisper, custom scripts.
Modality-independent architecture
Internal design is modality-agnostic. Shipped project types: images (detect / seg / classify), text, tabular, audio, video, LLM instruction/preference/RAG. Direction: documents/OCR, time series, 3D/point cloud, keypoints, OBB.
Reproducible
Every model should be traceable to dataset version, transform pipeline, schema, training config, source checksums, framework versions, weights, seed, and hardware.
Extensible
New capabilities arrive through plugins (oneopen.datasource, oneopen.trainer, oneopen.exporter, …) rather than hard-coded forks.