TinyBench trains compact sensor and audio models locally and exports an Arduino library. Start with a demonstration, then replace it with measurements from your own device.

1. Capture independent sessions. Upload the supplied capture sketch using Arduino IDE, close Serial Monitor, then connect over USB. Or import CSV/WAV; choose New audio project for microphone data.
2. Label and split. Use the same session ID for recordings from one collection session. Auto-split keeps those together. Capture at least three sessions per class; more diverse sessions give stronger evidence.
3. Train. Choose time statistics, sensor spectral features, or framed audio features. Classification, regression, and separately calibrated anomaly detection use training-only normalization.
4. Test once you are ready. Validation selects the training epoch. The separate test set is evaluated on request. Repeated tuning against test results makes it a validation set; collect new test sessions for a final assessment.
5. Export and verify. Install the ZIP library in Arduino IDE, run StaticBuffer, then integrate your sensor. Check memory and timing on the physical board.

USB capture requires a browser with Web Serial support, usually desktop Chrome or Edge, and HTTPS or localhost. No connection is requested until you click Connect. CSV and project import work without USB.

Autosave uses this browser’s IndexedDB. Browser cleanup can remove it. Save JSON backups to keep a portable copy. Offline reload becomes available after the app is cached on HTTPS or localhost.

This is an independent Green Shoe Garage tool, inspired by the data-to-device workflow. It is not an Edge Impulse product or a drop-in replacement.