What it is
A soundboard that reacts to a live call by itself, about a second after the moment.
What it does: Call audio is transcribed on a fixed 1.2 s clock, so text keeps flowing even when people talk over each other. Only the newest slice is judged. Jev answers a Noul (should a sound play now, following a free-text vibe) and a Choice (which clip, by its description). Code then applies the trigger level, the minimum gap and the no-repeat rule before anything plays.
Why it belongs in the showcase: A clear confidence-gated action loop in real time, with every rule outside the model and a live cost meter.
Caveats: Sends transcripts and audio to two providers; no published quality metrics.
Evidence
Public MIT README; speech-to-text (OpenAI Realtime by default) is the only piece that is not Jev; one Jev call per slice with `react_now` (Noul) and `clip` (Choice, matched against each clip's 'when to use it' description, or none); code owns the rules: trigger level (chill 88%, normal 75%, trigger-happy 60%), at least 4 s between clips, no repeat among the last 25; ships 15 synthesized sounds; author estimate about $0.30–0.40 per hour of constant talk for Jev plus transcription (prices as of October 2026); tests use a fake Jev and a fake transport.
Caveats
- Call transcripts go to TypeSafe and audio goes to OpenAI; the README asks you to tell people on the call
- The README demo is a written description, not a recorded benchmark; no reaction-quality metrics
- Per the README, Discord bots cannot yet reliably hear voice channels because of end-to-end voice encryption
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.