Assessment
AI Interview System
Gated assessment pipeline combining automated coding evaluation, AI avatar interviews with offline TTS, and live recruiter sessions — all monitored by OS-level integrity verification through the Tauri desktop client.
Assessment Pipeline
| Stage | Technology | Detail |
|---|---|---|
| Stage 1: Coding Assessment | Judge0 CE + Tauri | OS-level process monitoring. Detects AI copilots, remote desktop, browser automation. |
| Gate | Threshold Check | Coding score meets minimum threshold? Pass → Stage 2. Fail → rejected. |
| Stage 2: AI Avatar Interview | sherpa-onnx + Supertonic 3 | Offline TTS, Web Speech API STT, 3D avatar with lip-sync. Adaptive questioning via LLM. |
| Gate | Threshold Check | AI assessment meets threshold? Pass → Stage 3. Fail → rejected. |
| Stage 3: Live Recruiter | LiveKit SFU | Multi-participant WebRTC. 80-150ms audio, 120-250ms video latency. |
| Report | Integrity + Competency | Fused proctoring confidence + skill scores. Full evidence pack for recruiter. |
Supertonic 3 TTS Specifications
| Parameter | Value |
|---|---|
| Parameters | 99 million |
| ONNX Assets | 404 MB total |
| Languages | 31 |
| Runtime | CPU-only, no GPU required |
| Integration | sherpa-onnx Rust bindings |
| Privacy | Fully offline, no network dependency |
Speech Pipeline
| Component | Technology | Detail |
|---|---|---|
| VAD | Silero ONNX | Endpoint detection, streaming |
| STT | Web Speech API | Browser WebView native stack |
| LLM | Groq | Adaptive questioning, 350-600ms |
| TTS | Supertonic 3 ONNX INT8 | Offline, CPU, 31 languages |
| Playback | Audio Init | <50ms |
LiveKit SFU Performance
80-150ms
Audio Latency
120-250ms
Video Latency
<300ms
Multi-User Sync
<15%
CPU (4 participants)