Real-time AI interviews with BARS-anchored scoring, multi-modal cheat detection, and automated performance reports.

VU is a graduation project that conducts fully automated AI-driven job interviews. A candidate speaks to an AI interviewer in real time — the system transcribes their speech, evaluates their answers using BARS (Behaviorally Anchored Rating Scales), analyzes their voice for confidence, tracks their face and gaze via webcam, and detects cheating through 5 independent signals. At the end, the recruiter receives a complete performance report with sub-scores, per-question feedback, an overall summary, and an integrity label.

• Real-time voice interviews — dual WebSocket (STT + interview control) with edge-tts AI voice
• BARS-anchored scoring — 5 LLM-scored dimensions (1-5 scale with behavioral anchors)
• Multi-modal assessment — transcript analysis + audio confidence + video face/gaze tracking
• Cheat detection — 5 signals: tab switches, no face, multiple faces, gaze away, second speaker (diarization)
• CV analysis — 4 BARS dimensions scored from PDF/DOCX, skills extracted for question tailoring
• Adaptive questioning — difficulty-aware question count, topic diversity, LLM-driven follow-ups
• Pluggable LLM — switch between Gemini and Groq with one environment variable

Frontend (React)  ←→  AI Service (FastAPI)  ←→  Backend (NestJS)  ←→  PostgreSQL
                         ├─ STT (AssemblyAI)
                         ├─ LLM (Gemini / Groq)
                         ├─ TTS (edge-tts)
                         ├─ Video (OpenCV + YuNet)
                         └─ Scoring (BARS + cheat detection)

git clone --recursive https://github.com/vu-app-dev/vu-app.git
cd vu-app
cp .env.example .env   # fill in API keys
docker compose -f compose.yml -f compose.dev.yml up --build

→ Frontend at http://localhost:5173, Backend at :3000, AI at :8000

• Architecture — system design, request flow, data model
• AI Scoring — BARS dimensions, cheat detection, LLM config
• Deployment — VPS setup, DNS, SSL, backups

MIT