Try it — ComplaintGuard

Analyse your own transcript

The twelve cases on the front page are prepared, and they run without a model. This one does not: paste a call and the extraction and resolution layers actually run against it. Everything after that — the signals, the weights, the score, the intervention point — is the same deterministic code, so the answer is reproducible even though the reading of the words is not.

Please don't paste anything real. Text you submit here is sent to a model provider and held in memory on this server until it restarts. Everything else on this site is synthetic and stays that way — make something up, or use the example below.

One line per turn, each starting with customer: or agent:. Up to 60 lines and 6,000 characters. English only — extraction would cope with other languages, the signals and policy checks would not, and a confident analysis that found nothing is worse than a refusal.

This endpoint spends real model credits, so it is capped: 6 runs an hour per visitor and 80 a day for everyone. 80 left today. Identical text is answered from cache and costs nothing.