Agent tooling
rev trains a local decision model from labeled text and lets uncertain answers go to a person at a chosen cutoff.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The Python toolkit checks data, trains the model and calibrates scores, then serves decisions through a System One-shaped interface. This early project is separate from HCLTech's RevCT and TypeSafe's Jev. The README's performance comparisons remain the author's claims; they have not been independently reproduced.
Start with text for which you know the right decision. Before training, the data checks look for empty or conflicting examples. The toolkit uses the costs you assign to mistakes and human reviews to adjust when an answer should go to a person.