AI Interview System¶

HireMinds utilizes a highly advanced, low-latency AI interviewer.

Voice Pipeline Architecture¶

HireMinds supports two pipeline implementations behind an abstract interface (start, push_audio, manual_end):

Primary S2S WebRTC Pipeline: Self-hosted S2S WebRTC pipeline. Client-side VAD (~200ms latency), dynamic templated prompts, decoupled text/audio flow.
Legacy Fallback (V1): Older production pipeline over WebSocket. Fused STT+LLM+TTS with server-side VAD (2-4s latency).

Interview Lifecycle¶

sequenceDiagram
    participant C as Candidate
    participant H as HireMinds Backend
    participant D as Tevrix Voice AI 

    C->>H: Start Session Request
    H->>D: Initialize Context (Resume, GitHub, Questions)
    D-->>H: Session Token
    H-->>C: Connect Credentials
    C->>D: WebRTC/WS Audio Stream
    D->>D: Client-side/Server-side VAD
    D-->>C: TTS Audio Response
    D-->>H: Transcript & Events Stream
    H->>H: Persist Turn & Enforce Budget
    D-->>H: Session End Webhook
    H->>H: Scoring Service Grades Transcript

AI Scoring¶

Post-interview, the ScoringService evaluates the full transcript (from DB) against a rubric.

Scores include star_compliance, confidence_signal, and technical accuracy.
System failures (e.g., <4 turns, crashes) return None (manual review flag) to ensure fair grading (no false negatives).
Partial transcripts are scored by the session_watchdog if stuck for >30m.