Fast, low-cost real-time agents.

A voice agent makes a dozen small decisions every turn: is this a card number, do they want a human, is my draft safe to say? jevelin takes them off the critical path with a System One classifier, either Jev or the Apache-2.0 Laya on hardware you already own, and schedules every turn as a dependency graph. Move the sliders. The Python engine is running in your browser.

Loading Python in your browser (first visit downloads about 10 MB)…

Assumptions: drag to change them

Same callers for both pipelines. Illustrative latencies, not production measurements.

When does the agent start speaking?

Fast pathBaseline: LLM judge in series

Share of turns that have started speaking by each moment after the caller stops. Further left is faster. Dots mark each median. Hover to read values.

Baseline vs fast path

One turn, stage by stage

Time zero is the moment the caller stops talking. Everything left of it happened while they were still speaking.

Classifier (Jev) LLM Speech out Rules + router Caller speaking Discarded guess Timed out

Baseline

Fast path

The design rule

Raise the drawbridge. Never lower it.

A classifier can make a turn more cautious. It can never make it less cautious. Every output leads to a stricter, pre-approved path: a protective script, a transfer, a blocked draft, a hold. None of them unlocks anything.

  • Rules beat the classifier. After hours, "transfer me" becomes a callback, whatever the model says.
  • Partial transcripts add caution but never route. A card number is caught before the caller finishes reading it.
  • Fail closed. A timeout gets the hold script. A missing answer never counts as a pass.

How the fast path works

  • Classify while they're still talking. One classifier call answers every question in parallel.
  • Script the routine turns. Balance, due date, outage: pre-approved text, no LLM call.
  • Send the rest to the smallest model that fits. When the router is unsure, it uses the large model.
  • Generate early. Start on the partial transcript's guess, and throw it away if the final transcript disagrees.
  • Guard every generated draft. Nothing unchecked is spoken.

What's simulated: LLMs, speech and the classifier's answers, with latencies and prices from a JSON profile you can replace. What's real: the scheduling, routing, rules and fail-closed logic. It's the same code as the repo, with tests.