XERPIUM AI INTERVIEW — AI Recruitment

Summary

AI Recruitment Ecosystem

XERPIUM AI INTERVIEW is an AI-based recruitment solution — covering automatic CV screening (score 0-15, bias redaction, recommendations), adaptive AI interview (backend-governed, difficulty scoring, 9-level recommendation, risk flags), and a self-service candidate portal.

📝

CV SCREEN

Automatic & Unbiased

🎤

INTERVIEW

Adaptive AI Engine

📊

SCORING

9-Level Recomendation

🌐

PORTAL

Candidate Self-Service

CV Screening

AI CV Screening Engine

📝 Screening Pipeline

Pipeline — Extract → Normalize → Strip Bias → Token Safety → LLM → Parse → Decide → Save
6-State Analysis — Pending → Processing → Completed / Partial / API Failed / Extraction Failed
File Support — PDF via pdfplumber (primary), PyPDF2 (fallback); DOCX via python-docx; .doc via latin-1
Token Limit — 4000 character safety limit per section
Retry — Max 3 retries per LLM call; safe retry with cached extracted text (no re-extraction)
Partial Result — Missing JSON keys filled from defaults, never drops valid fields

🔎 Scoring & Recommendation

3 Dimensions — Experience (0-5), Skill (0-5), Education (0-5) = Total (0-15)
Confidence Index — 0.0 - 1.0 with label: High Fit, Moderate Fit, Low Fit
Recommendation — Recommended (≥12), Review (≥8), Reject (<8)
Backend Override — System forces recommendation based on score, not just LLM output
Summary — Executive summary, strengths, weaknesses, HR insight, interview focus area
Auto Interview — Button to auto-create interview invitation for recommended/review candidates

🔒 Bias Redaction

Auto-Stripped — Email, phone number, gender, religion, age, photo — all removed before the LLM
Section-Aware — Token limit applied per CV section, not in total
Structured Log — JSON operation log for audit
Partial Mode — If the LLM partially fails, the backend continues with valid data
Retry Analysis — Re-analysis via action button (safe — uses cached text)
Bias-Free Scoring — Pure score based on competence and experience

📋 Applicant Integration

hr.applicant — Extended with AI fields: ai_cv_score, ai_cv_recommendation, ai_cv_confidence
Processed Flag — ai_cv_processed: processed or not
Auto Button — action_process_ai_cv_review() — one-click CV processing
AI Summary — Executive summary, strengths/weaknesses, HR insight
Interview Focus — ai_interview_focus_area — interview topic recommendations
Status — Processing status per applicant, real-time feedback

AI Interview

Adaptive AI Interview Engine

🎤 Interview Engine

Adaptive Loop — Backend-governed: LLM = text generation only; Backend = ALL decisions
State Machine — Pending → Running → Evaluating → Done — full lifecycle
Difficulty Scoring — Easy (0.8x), Medium (1.0x), Hard (1.2x) multiplier
Intent Detection — Relevant, Partially Relevant, Irrelevant, Clarification Request, Off-Topic
Controlled Retry — Max 3 retries per question
Escalating Hints — 3 levels: Soft → Structured → Final — hints increase gradually
Off-Topic — Max 3 irrelevant answers → question skipped

📈 Scoring & Evaluation

Per Question Score — Objective progress + question scores + difficulty multiplier
Critical Gate — If a critical question fails, the recommendation is capped at 'borderline'
Risk Flags — 5 flags: fundamental_knowledge_gap, failed_easy_question, high_hint_dependency, multiple_off_topic, system_evaluation_error
9-Level Recommendation — 1-9 scale + harmonization + risk downgrade
Evaluation Status — Full / Partial / Failed — transparent
Heuristic Fallback — Alternative scoring when the LLM fails
Performance Tier — Final tier after harmonization and risk adjustment

📜 Question Bank & Survey

Survey Integration — survey.survey extended: ai_active, ai_mode (fixed/adaptive), ai_passing_score
Per Page — ai_weight, ai_min_score, ai_description
Per Question — ai_difficulty (easy/medium/hard), ai_critical_flag, ai_followup_hint, ai_evaluation_hint
Weights — ai_question_weight (default 5.0), ai_question_min_score (default 2.0)
Topic Locking — Backend decides topic order, LLM never skips
Language — Indonesian & English with language lock — prevents switching

🌐 Candidate Portal

Portal URL — /ai-interview/{access_token} — secure per-candidate link
Invitation — Draft → Sent → Started → Done → Expired (7-day default)
Auto Invitation — Sent automatically from CV screening for recommended candidates
Early Termination Protection — LLM cannot generate closing statements; backend has full control
Output Sanitization — Forbidden opening phrases blocked/replaced
Flow Integrity — Flow integrity failure logging for debugging

Comparison

Why XERPIUM AI INTERVIEW?

Aspect	Manual / Standard ATS	XERPIUM AI INTERVIEW
CV Screening	HR baca manual, bias	✓ Automatic AI, bias redaction, 0-15 scoring
Bias	Gender, age, religion	✓ Auto-stripped, pure competence
Interview	Manual, inconsistent	✓ Adaptive AI, difficulty scoring, risk flags
Scoring	Subjektif	✓ 9-level, backend-governed, harmonized
Portal	✗ None	✓ 24/7 self-service, token-based access
Integration	Separate ATS	✓ FULL XERPIUM: applicant, survey, recruitment