MenuClose
PL
← All projects

Pilot · 2026

ColdCallAI

A voice agent calls leads and a model extracts facts from the call — the score is computed by code, not AI.

  • AI / LLM
  • Business apps
Abstract dot field on a yellow background

A cold-calling system for real estate: a voice agent runs the call, a language model extracts facts backed by quotes from the transcript, and deterministic scoring from a config file rates the lead.

Problem

Rating leads after a call is subjective and cannot be repeated. When a language model gives the rating, the same lead can get different scores, and nobody knows why.

Solution

Separate the roles: the model only extracts facts, and each one must be backed by a verbatim quote from the call or it is discarded. The score is computed by simple, versioned code from YAML rules, so the result is repeatable and explainable.

  1. Voice agent calls
  2. Transcript via webhook
  3. Fact extraction with quotes
  4. Scoring from YAML rules
  5. Lead in the panel
From a call to a rated lead
of facts backed by a quote from the call
100%
scoring decisions made by the model
0
requirement specifications
11

How it is checked

A replay tool runs recorded calls through extraction and scoring again after every change, and a golden sample set shows whether results have drifted. An outbound dialer, a panel for reviewing calls and a watchdog for shared hosting complete the system.

Related projects