nodd

Tiny, calibrated classifiers for one decision each.
Small enough to run anywhere, even in the browser, and they know when they are unsure.

Every model here is a small MiniLM encoder (about 24 MB) fine-tuned for a single task, such as moderating comments, spotting prompt injection, reading sentiment or routing support messages. Each prediction comes with a calibrated confidence: when the model is sure it decides on its own, and when it is not, you can hand the input to a bigger model.

The same model runs wherever you need it: in Python with PyTorch or ONNX Runtime, on a Node server, or fully offline in the browser with no per-call cost. Exports are checked for parity, so every runtime gives the same answers.

The models are trained on synthetic data, so expect somewhat lower accuracy on real traffic. Each model page lists its labels, metrics and how to use it.

GitHub · Live demo · Docs