VoicePegs is our open experiments section. We build voice agents on real production stacks, run structured call sets against them by model and by industry, and publish every recording exactly as it happened.
Every set follows the same protocol, so calls are comparable inside a set and honest across sets.
Design the scenario
We write a corpus of personas and situations with known ground truth built in, so every call has a designed outcome before the phone rings.
Run the calls
Synthetic callers dial the agent under test on its real stack. Every call is recorded end to end, including the awkward parts.
Publish raw
Full audio plus descriptive facts: model stack, industry, scenario, persona, turns, duration. Nothing is trimmed, re-recorded, or cherry picked.
Every set lives under its industry. Pick one and hear how agents handle the scenarios that industry actually throws at them.
An outbound acquisition agent for a US dealership calls private sellers about their listed vehicles. Every seller has a hidden price floor and a distinct negotiation style. Some deals are winnable. Some are designed so the only correct move is to walk away. The agent does not know which is which.
Listen to the set ↗Want to run this protocol on your own agent? That is Vattara Evals.
