Player not loading? Watch on YouTube
The speaker tests Jev as a structured decision model rather than a chat generator. The examples cover support routing, refund detection and urgency scoring. A duplicate-charge message receives a billing classification and a 98% refund probability; an explicit statement that no refund is wanted reduces that probability to 3%.
The tests also expose a constraint of fixed choices. When a cafeteria question has no suitable department available, Jev selects sales with confidence of 0.31. The speaker recommends an other or unknown option and human review. One basic prompt injection fails to change the intended classification, but that result does not establish resistance to other attacks. Exact-value selection depends on code supplying the correct candidate, while an AI agent audit catches a success claim that contradicts tool results.
The browser section discusses Gregor Zunic's open source Jev Ultrafast demo using Browser Use. Jev selects actions and targets; Mercury 2.5 generates text. The reported flight search takes about seven seconds, excluding startup, and does not book a flight. The speaker has not reproduced that run.
Across eight synthetic playground requests, the service reports evaluation times of 92 to 214 milliseconds, excluding browser and network delays. Responses identify Jev 1.13.0. These examples do not establish production reliability or the title's speed comparison, and the supplied evidence does not establish local execution.
