Jevstiller is an open-source tool that distills Jev's outputs into a local model. Jev itself is a fast, cheap AI that answers only structured question types like Choice, Score, and Noul. Jevstiller acts as a cache in front of Jev, learning which requests a user typically sends and answering those locally, while forwarding anything unfamiliar to Jev's servers.
The tool starts by passing all queries to Jev to collect training data. Once enough samples exist, it fits a student model and a routing policy. Confident, familiar inputs are answered locally; everything else goes to Jev unchanged, and each response becomes a new training row. According to the developers, local answers can come back in as little as 15 ms from a device's CPU, with no Jev tokens spent.
To keep the local model in step with Jev, Jevstiller always sends a fixed 2% of requests to Jev as an audit. If agreement falls below a configured target, the audit rate rises and retraining accelerates. If agreement collapses entirely, all traffic returns to Jev and training restarts. In a 24-hour test, a stand-in Jev silently changed every answer at hour twelve; the local share dropped from 90% to 9% within four minutes and recovered to 90% within 49 minutes.
The developers caution that agreement is not accuracy. Jevstiller only mimics Jev's behavior, including its mistakes, so a locally answered query is not necessarily more correct than one sent upstream.