Classifying 1,500 Real Emails

The post does not say. It gives a corpus of 1,500 emails from one inbox and calls the result blown away, with no label set, no human ground truth, no error count, and no baseline classifier to compare against.

A demo. It is the author's own mail with no methodology published, so it cannot tell you how the model handles your labels, your languages, or an awkward edge case. It is a useful signal that the task shape works, and it is not an accuracy number.

A reply raises exactly this: classifying emails one at a time burns round trips that batching would amortize. The post does not answer, so if your mailbox is large, requests per second and cost per thousand emails are the two things to measure yourself.

Not stated. The post quotes a corpus size and a reaction, but no spend and no wall-clock time. A separate account reports 500 emails classified in seconds for a few cents, which puts the order of magnitude in view, though this post itself gives nothing to check against.

None is stated. He is a developer posting demos of the model and telling people to follow for more Jev demos, and he also links a separate video demo of his own. That makes the 1,500-email run an enthusiast's test rather than a commissioned evaluation.

Not from the post. The result is shared as a video, with no repository, no prompt, no label taxonomy and no harness published, so there is nothing to rerun. Reproducing it means building the classifier around the API yourself.

Yes, a second demo video. After the email test he posted another Jev demo titled Meet Jev: Future of AI and asked followers to stay tuned for more, so the email run is one in a short series of launch-week demo posts rather than a standalone study.