Jev Decision Model: Forecasting AI Receptionist Calls from Structure Alone

Melvin Vivas · X video post · 2026-09-30 · 0:11 · 861 views · Open on X

Topics: AI Agents, Tool Use & MCP, LLMOps, Deployment & Monitoring, Industry Trends & Job Market · Level: intermediate

Summary

Melvin Vivas shares a real-world use case of Jev, which the post calls a "decision model." In the quoted post, Jev was given 2,029 real phone calls from an AI receptionist. It had no transcripts or audio, only the structure of each call: turns, tool calls, workflow stages and timing. It made turn-level forecasts quickly enough to run during live calls. The 11-second video's machine transcript is unusable, so this note is based only on the caption.

Key points

• Jev is described as a 'decision model' used alongside a voice-agent (AI receptionist) system.
• Dataset: 2,029 real phone calls handled by an AI receptionist.
• No transcripts or audio were used. Calls were reduced to structure only: conversation turns, tool calls, workflow stages and timing.
• Jev made 38,012 turn-level forecasts across those calls, about 19 per call.
• Median forecast latency was 118 ms, fast enough to run in real time during live calls.
• Takeaway: agent metadata such as tool-call traces, workflow state and timing can be enough to predict call behavior, and it avoids handling sensitive conversation content.
• The caption is cut off, so the quoted post doesn't show what Jev was predicting (for example, outcomes or escalations) or how accurate it was.

Resources mentioned

• Jev · tool · typesafe.ai · paid
  A decision model that makes turn-level forecasts on AI agent conversations using only structural signals (turns, tool calls, workflow stages, timing).
  Also in: Generating a Repo Promo Video with a Claude Skill on Sonnet 5.5 vs Opus 5.5 (Melvin Vivas on X · notes), GLiDE by Fastino Labs: A Post-Trainable Reasoning Decision Model (Melvin Vivas on X · notes), Using Jev as a Reranker to Augment RAG Retrieval (Melvin Vivas on X · notes), Using Jev (TypeSafe AI) as a Decision Model for Email Classification (Melvin Vivas on X · notes) and 12 more

Try this

• Log structural traces from a voice or chat agent (turns, tool calls, workflow stage, timing) and train a lightweight model to forecast call outcomes turn by turn without using transcript content.
• Measure the latency of a real-time per-turn prediction model and aim for a median around 100 ms so it can run during live calls.