Jev Daily

Kunal Abichandani's voice coffee order renders in about 200ms

A single demo, a single number, and no breakdown of where the latency goes.

Kunal Abichandani built a voice coffee ordering system on Jev in which the drink, and any changes to it, appear on screen while he is still speaking. He puts the gap at roughly 200ms from word to screen. The post is a video demo with that one figure attached.

Kunal Abichandani
@kunalabcdani
X
The drink and any updates show up while I’m still mid-sentence, ~200 ms from word to screen.
Nov 9, 2026 · View on X

That is the whole of the evidence. He does not say how the 200ms was measured, whether it is a median over a run or a number he watched once, or where the clock starts and stops. He does not say what sits in front of Jev on the audio path, which speech-to-text service is doing the transcription, or how much of the budget that step eats. He does not describe the menu either, so there is no way to tell whether the model is picking from a dozen drinks and a handful of modifiers or something wider.

What the number does and does not tell you

A voice interface that updates mid-sentence is a different interaction from one that waits for you to stop talking, and the difference is mostly latency. So the figure is the interesting part of the post. It is also one builder's figure on one flow he built himself, which makes it a demo and not a benchmark.

For people building on it, the thing to be careful about is attributing the 200ms to Jev. A voice pipeline has at least a transcription hop and a render step either side of the model call, and nothing in the post splits the budget between them. If you are sizing a similar build, the number to get is your own, measured per stage, before you decide the model call is the cheap part or the expensive one.

There is also no cost figure here, no error rate, and no account of what happens when the order is ambiguous or the speaker changes their mind twice. Abichandani's summary of the experience is that it "feels pretty magical", which is a demo verdict, not a reliability one.

Know what Jev is good at before you bet on it

One email a day. Under three minutes. What builders found out about Jev yesterday.

Get the next one by email

Jev, read daily so you do not have to. The builds, the benchmarks, the criteria that worked and the cases where it lost, from the people shipping on TypeSafe AI's System One model.