Voxeval

How it works

From business context to a release decision.

One continuous evaluation flow connects the calls you run to the evidence that gates production.

  1. 01

    Import business context

    Add workflows, policies, tool requirements, IVR maps, call transcripts, and prior incidents.

    • workflow/scheduling
    • policy/health-12
    • tool/schedule.create
  2. 02

    Generate domain-specific evals

    Define expected actions, required disclosures, allowed tool calls, caller personas, and success criteria.

    • 128 call scenarios
    • 8 caller personas
    • 5 audio conditions
  3. 03

    Run your actual voice agents

    Exercise the same workflow across accents, noise, low volume, interruptions, silence, and IVR paths.

    • noisy mobile
    • interruptive caller
    • low volume
  4. 04

    Evaluate the outcome

    Inspect audio, transcript, tool sequence, workflow state, policy adherence, latency, and final result.

    • identity verified
    • tool call correct
    • disclosure missing
  5. 05

    Gate the release

    Return Ready, Caution, Blocked, or Not Enough Data with linked evidence and reusable regressions.

    • decision: blocked
    • regression: generated
    • CI exit: 1
Illustrative evaluationAppointment scheduling
Call in progress
Caller · 00:18

“Can you move my appointment to Friday afternoon?”

Agent listening
Evidence stream3 of 4 checks complete
  1. 01
    Identity verifiedDate of birth matched
    Pass
  2. 02
    Tool call correctschedule.update()
    Pass
  3. 03
    Policy disclosureConsent language missing
    Review
Release decisionBlocked by one policy requirement
View evidence
Illustrative evaluation data showing how call evidence informs a workflow-level release decision.

Release gate

Evidence becomes a decision your pipeline can enforce.

Every blocked release points back to the policy, call, and regression that produced the decision.

$ voxeval gate release/v2.3.0-rc1
decision: blocked
blocker: policy/health-12
regression: generated
exit: 1