CONTROLAgents & automation
Jev Decision Layer for Voice Pipelines
Himanshu@Fabulous_7781
2:00Jev acts as a decision layer in voice pipelines, utilizing typed primitives like Noul (yes/no with probability), Choice (labeled decision), and Score (scalar on a defined rubric) to process conversational turns before and during LLM responses.
SignalThree typed primitives, two calls per turn: Noul — a yes/no with a probability Choice — a labelled decision with your own criteria Score — a scalar on a rubric you define Call 1 runs before the LLM.
I used Jev (@typesafeai ) as the decision layer in a Pipecat voice pipeline. Three typed primitives, two calls per turn:
Noul — a yes/no with a probability Choice — a labelled decision with your own criteria Score — a scalar on a rubric you define
Call 1 runs before the LLM. Noul("is this a complete thought?") ends the turn semantically instead of on a silence timer, and Choice("which support flow?") rides along in the same request, so intent reaches the LLM on the turn it belongs to. Batching that second question cost 6 ms.
Call 2 runs while the LLM is already writing, so the caller never waits on it. Score for frustration on a hostile→delighted rubric. Noul for sarcasm, repetition, confusion, satisfaction. Choice for what they actually need — acknowledgement, a concrete fix, information, or a human. Six questions, one call.
The sarcasm check is the one that earns its place. On that line the tone-contradiction Noul hit 0.92.
Jev's state is text-only, so pitch, pace and volume get bucketed into phrases like "rapid and clipped" relative to that speaker's own baseline, and the model fuses delivery with the words rather than scoring them separately.
Handoff fired on turn 5. The caller didn't ask to cancel until turn 8.