Make real AI reliable: raise timeouts (45s) + max_tokens for gemini-2.5 thinking so scoring/agent/CV calls return JSON instead of timing out to fallback; remove public test-email endpoint

Use current Gemini model via GEMINI_MODEL env var (default gemini-2.5-flash); gemini-2.0-flash is deprecated/shut down

Agentic AI (#4): autonomous interview agent decides probe-vs-advance after each answer, generates targeted follow-ups, logs decisions to the report

Emotion/sentiment analysis: derive facial emotion from blendshapes, add emotion_analysis (dominant/distribution/composure) to report

Don't score transcription failures: null confidence/clarity/engagement when audio can't be processed; backfill repairs existing reports

Add backfill endpoint to link existing completed interviews to applications

Link interview score + session to application so the company dashboard shows results

Report: full violation objects in integrity_flags, skill_breakdown as dict, robust candidate-name lookup

Eye tracking: EAR blink detection, attention/suspicion scores, feed integrity; live HUD in interview

Interview: hang-proof AI calls (thread timeout), JSON structured scoring, skill breakdown + quoted report

Interview: fail-fast on Gemini quota/429 (max_retries=1) to avoid request hangs

Interview upgrades: MongoDB persistence, adaptive follow-ups, voice analysis, integrity score

Initial commit: BotBoss backend (own deployment)