InterviewAI MVP：48 小时构建语音交互模拟面试

MVP Architecture Under Pressure

InterviewAI chains four ideas: understand the candidate's background, tailor questions to role and skill level, conduct a voice conversation, and score against clear dimensions. No fixed script — each session adapts.

We shipped in 48 hours by scoping ruthlessly: one interview format, one voice flow, one scoring framework. Band9AI later inherited the evaluation discipline from this build.

Session Flow (Conceptual)

{
  "interview_flow": {
    "inputs": ["resume", "target_role"],
    "phases": ["warmup", "technical_depth", "problem_solving", "behavioral"],
    "scored_dimensions": ["reasoning", "communication", "relevance"],
    "output": "structured_scorecard_with_action_items"
  }
}

Lessons for Band9AI Speaking

Pressure-faithful simulation requires fast feedback and visible scoring dimensions — the same principles behind IELTS speaking mocks. InterviewAI proved the loop; Band9AI scaled the evaluation mindset.

Try the MVP at interviewai.ca.

What We Kept Despite Speed

Shipping in 48 hours does not excuse skipping evaluation. InterviewAI still needed structured outputs, timing fidelity, and clear practice labeling. Speed came from narrowing scope — one interview loop — not from deleting the reliability layer.

That tradeoff is repeatable: constrain the surface area, keep validators, and measure against a tiny golden set before inviting real users into the loop.