Your phone trains the model — but only when it's cool and plugged in

What it does

FractalAndroid is the mobile client of the Fractal distributed compute framework. It turns an Android phone into an edge node that runs on-device federated training via TFLite — downloading model slices and data bins from the central FractalCore server, training locally, and uploading encrypted checkpoints back. Raw user data never leaves the phone; only trained weights do.

The interesting bit

The OperationControl layer is the star: before and during every training round it samples battery SoC, charging status, SoC temperature, and memory pressure, and pauses or throttles work when thresholds are violated (e.g. battery below 50% and not charging, or thermals above 40°C). The README calls this “resource empathy” — it’s essentially a hardware governor bolted onto a training loop, which is exactly the piece most volunteer-computing apps skip.

Key highlights

• On-device TFLite training with Image_Trainer, DataManager (ByteBuffer batch loading), and CheckpointManager for .ckpt serialization
• Multi-variable telemetry gating: battery, thermal, unmetered-Wi-Fi, and memory checks pause, defer, or throttle execution
• MVVM architecture with clean separation between UI, gating, execution, and network layers
• “WiFi Bank”: validated checkpoint uploads credit liquid MBs to a Firestore ledger, redeemable for bandwidth — compute-for-data barter
• TLS-encrypted weight egress via ModelTransmitter; task polling over REST HTTPS

Caveats

• Proprietary Non-Commercial v3.0 license — you can read it, not ship it
• Depends on the closed FractalCore server and Firebase/Firestore; this repo is one half of a system, not a standalone framework
• The reward economics (bandwidth credits) are described only at the UI level; no validation or anti-cheat detail is given

Verdict

Worth a look if you’re building mobile federated learning and want a reference for telemetry-gated execution — the gating design is genuinely instructive. If you wanted an open, self-hostable FL framework, this is a client for someone else’s network.