Moritz Kremb built a Jev sales copilot that scores closing odds
The demo runs on a recorded call, and no latency, cost or accuracy numbers came with it.
Moritz Kremb built a sales copilot on Jev that listens to a sales call, tells the rep what to say next, helps them follow the script and handle objections, flags which stage of the call they are in, and shows a live probability of closing. He posted a demo. The demo runs on a recorded sales call.
Gives you live signals and probability of closing
That last detail is the one that matters. A recording is a fixed audio file with a known end, which means the classifier can be fed clean, complete chunks at whatever pace the harness chooses. A live call is duplex, people talk over each other, the audio arrives in fragments, and the stage label has to be right while the rep is still in the sentence it is supposed to help with. Kremb's post describes the first case and claims the second as the use case.
What was not posted
No latency figure. No cost per call. No accuracy number for the stage detection, and no calibration for the closing probability, which is the one output here that is a number and therefore the one most likely to be read as authoritative on a screen during a live call. There is no indication in the post of how many calls the demo covered beyond the one shown, or what the script and objection handling were compared against.
So what exists is a working build and a video, not a measurement. That is a normal place for a demo to be on the day it ships, and it is worth seeing what the interface looks like when a fast classifier sits underneath a real time UI rather than behind an API call.
For people building on it
The two claims in this build fail differently. Stage detection is a classification problem where you can eventually count how often it was right. A live closing probability is a calibration problem, and a number that is confidently wrong at 78 percent is worse than no number at all, because the rep will change what they say because of it. Neither has been evaluated here. Treat the demo as evidence that the pipeline runs, not that the outputs are trustworthy mid call.