Sharing my final project: Forsa AI, a voice-first interview practice platform.
What it does: Candidate picks a role (any role, even a custom title) and optionally uploads a CV → has a real live voice interview with an AI interviewer (GPT-4o via Vapi, WebRTC) → gets an instant scorecard from Claude, cross-referenced against their CV, plus a downloadable PDF action plan with free courses matched to their weakest areas.
A few engineering decisions I'm happy with:
• Two-model architecture: GPT-4o live for speed, Claude post-call for structured, schema-guaranteed scoring
• "Smart Mute" — muting mid-answer tells the agent to wait, not skip the question
• A 30-case automated eval suite (npm run eval) that verifies the real prompt/logic wiring, not mocks — catches real regressions like a dead course link before submission, not after
Thanks to everyone here for the feedback along the way — happy to answer questions about the Vapi/Claude wiring if anyone's building something similar