Our voice interview platform
The Voice Never Grades
The voice agent that conducts an interview is built to be friendly. Which is exactly why it is not allowed to judge: good listeners make lenient judges.
Scoring falls to a second model that assesses each answer on its own — criterion by criterion, with supporting passages from the conversation. And even that verdict is only a proposal: the final word belongs to deterministic PHP code.
- Division of roles
- speaking and scoring kept separate
- Every verdict
- evidence cited per criterion
- Final word
- deterministic code
The Model Proposes, the Code Disposes
After every answer, the scoring model delivers a typed, schema-validated finding: criterion by criterion, each backed by the passages in the transcript that support the verdict. No prose, no gut feeling — structured output that code can check.
The decision itself is then made in PHP: rules that are transparent and testable. If the model fails or returns something unusable, the system degrades in a controlled way instead of guessing. And when a human disagrees, they simply overrule the system — that is by design, not an exception.
Every scoring prompt is versioned and traceable by hash: for any verdict, you can say which version produced it. The decisions even survive deletion of the transcript — deliberately, because accountability and data minimization have to work together.
Areas
More from the AI Spotlight
Back to overview