Bjorn Cortex

A centralized training server for the Bjorn swarm. Each Bjorn device explores its environment and sends anonymized feedback to the Cortex. The Cortex aggregates field experience across the fleet, trains a model, and redistributes it to all connected devices.

The current version targets local experimentation and iterative testing. Production deployment on a VPS is a future milestone.

---

How it works

1. Bjorn devices send anonymized .csv.gz training files to the Cortex upload endpoint.
2. The Cortex trains a TensorFlow/Keras model on the aggregated data.
3. The latest model is served back to connected devices on request.
4. Live training status is streamed via WebSocket.

---

Features

• FastAPI backend and dashboard UI.
• Upload endpoint for .csv.gz training data.
• Training pipeline with TensorFlow/Keras.
• Live status via WebSocket.
• JWT authentication with MFA (TOTP).

---

Requirements

• Python 3.10+
• TensorFlow-compatible environment (GPU optional)

pip install -r Cortex/requirements.txt

---

Running

cd Cortex
python server.py

Access the dashboard on http://localhost:8000.

---

Docker

docker compose up -d --build

---

Security

Local configs (security_config.json, server_config.json) can contain secrets and are git-ignored. Commit only the sanitized .example variants. Rotate any secret previously exposed before pushing.

Do not expose this service publicly without HTTPS and a reverse proxy.

---

API endpoints

Method	Path	Description
GET	/setup	Initial setup page
GET	/setup/qr	TOTP provisioning QR
POST	/token	Issue JWT (OAuth2 password flow)
GET	/	Dashboard
POST	/upload	Upload training data
POST	/train/start	Trigger training
GET	/model/latest	Latest model metadata
GET	/model/download/{filename}	Download model artifact
WS	/ws/logs	Real-time training events

---

Star History

---

License

MIT.