curiositech/interview-simulator

Interview Simulator

Platform architecture and coaching system for realistic mock interview practice with adaptive difficulty, emotion-sensitive voice AI, and comprehensive performance tracking.

Decision Points

Difficulty Adjustment During Session

IF emotion_data.hesitation_rate > 70% FOR 2+ consecutive responses
├── AND current_difficulty >= 3
│   └── REDUCE difficulty by 1 level
│       └── Voice persona: "Let's back up to fundamentals"
│
├── ELSE current_difficulty < 5
│   └── MAINTAIN difficulty
│       └── Voice persona: "Take your time, think through this"
│
└── ELSE (hesitation at difficulty 1-2)
    └── PAUSE session for confidence building
        └── Offer: "Would you like a 2-minute break?"

IF technical_accuracy > 85% AND response_time < 60% of target
├── AND current_difficulty < 5
│   └── INCREASE difficulty by 1 level
│       └── Add follow-up: "Now optimize for X constraint"
│
└── AND current_difficulty = 5
    └── MAINTAIN but add adversarial prompts
        └── "What if the requirements changed to..."

Round Type Selection

IF any dimension_score < 60 FOR last 3 sessions
├── SELECT lowest_scoring_round_type
│   └── Override user preference if different
│
ELIF no round_type practiced in last 7 days
├── SELECT oldest_unpracticed_round
│   └── Variety bonus for skills retention
│
ELSE
├── SELECT weakness_weighted_random()
│   └── 70% bias toward bottom 2 scores, 30% random

Proctor Flag Response

IF gaze_off_screen > 5 seconds
├── IMMEDIATE: Voice AI pauses, asks "Everything okay?"
│   └── IF happens again within 10 minutes
│       └── FLAG session as "distracted practice"
│       └── Debrief includes focus recommendations
│
IF second_monitor_detection = true
├── IMMEDIATE: Session pause, warning overlay
│   └── "Real interview won't have external resources"
│   └── Options: [Continue with flag] [Reset clean]
│
IF note_taking_detected (hand tracking)
├── ALLOW in Training mode with flag
├── BLOCK in Simulation mode
    └── "Please close notes for realistic practice"

Failure Modes

1. Overtraining Comfort Zone (Behavioral Looping)

Symptoms: >50% of sessions are same round type, plateau in target skill, false confidence in demo sessions Detection Rule: IF session_history.last_10_sessions contains >7 of same round_type THEN flag "comfort_zone_looping" Fix: Lock out preferred round type for 72 hours, force weakness-weighted selection, show heatmap of practice distribution

2. Emotion Detection False Positives (Technical Hesitation)

Symptoms: Difficulty drops during complex technical problems where thinking time is normal, candidate reports sessions feel too easy Detection Rule: IF difficulty_reduction triggered AND problem_complexity = high AND response_quality > 75% THEN flag "false_positive_hesitation" Fix: Calibrate hesitation baseline per round type, extend thinking time allowance for design problems from 30s to 90s, separate "thinking" from "struggling" in emotion model

3. Network Latency Ghost Sessions

Symptoms: Voice AI cuts out mid-question, session scoring incomplete, debrief missing transcript sections, user reports poor experience Detection Rule: IF voice_stream_interruptions > 3 OR transcript_coverage < 80% THEN flag "network_degraded_session" Fix: Implement local recording fallback, retry logic for API calls, session quality gate before scoring (reject if <80% data integrity), offline mode with sync-later

4. Proctor Calibration Drift

Symptoms: False flags for normal behavior, user disables proctoring due to oversensitivity, gaze tracking works poorly for specific users (glasses, lighting) Detection Rule: IF proctor_flag_rate > 0.3 per session AND user_session_count > 5 THEN flag "proctor_oversensitive" Fix: Per-user calibration phase (30-second normal gaze pattern recording), lighting condition adaptation, glasses detection with adjusted thresholds

5. Debrief Analysis Inflation

Symptoms: All session scores >80%, improvement suggestions too generic, user reports disconnect between debrief and actual performance Detection Rule: IF average_session_score > 80 AND session_count > 10 AND difficulty_level unchanged THEN flag "score_inflation" Fix: Recalibrate scoring rubric with harder baseline, cross-validate with round-specific skill scoring, add "compared to senior candidates" percentile context

Worked Examples

Example 1: ML Design Session with Difficulty Adjustment

Setup: User has 3 previous ML design sessions, scores: [65, 72, 68]. Starting difficulty: 3/5.

Session Flow:

00:00 Voice AI: "Let's design a recommendation system for a video platform. Start with the problem setup."
00:45 User: "Um... so we need to recommend videos... I guess we'd use collaborative filtering?"
      Emotion data: hesitation_rate=85%, confidence=30%
01:15 Orchestrator detects: high hesitation + low confidence = reduce difficulty
      Difficulty: 3 → 2, Voice persona shift to "supportive"
01:20 Voice AI: "Great start! Let's break this down step by step. What data do we have about users?"
02:30 User: [More confident response about user features]
      Emotion data: hesitation_rate=40%, confidence=70%
      Orchestrator: maintain difficulty=2, user finding footing
15:00 User demonstrates solid understanding at level 2
      Technical accuracy=80%, response time=good
15:30 Orchestrator: increase difficulty 2 → 3
15:45 Voice AI: "Now, how would you handle the cold start problem for new users?"
30:00 Session ends, user handled level 3 questions well

Debrief Output:

• Started at difficulty 3, adjusted to 2 due to initial hesitation, finished at 3
• Strength: Good recovery and adaptation to feedback
• Weakness: Initial problem setup needs practice (recommend system design fundamentals)
• Next session: Start at difficulty 3, focus on problem scoping

Example 2: Network Latency Recovery

Setup: User in poor connectivity environment, voice stream unstable.

Session Flow:

00:00 Session starts normally
02:15 Voice stream cuts out mid-question
02:20 Orchestrator detects stream_interruption, activates fallback
02:25 Text overlay: "Voice connection unstable. Switching to text mode."
      Question continues in text, user types responses
15:30 Voice reconnects, quality test passed
15:35 Orchestrator: "Voice is back. Continuing with audio."
30:00 Session completes

Quality Gate Check:

• Transcript coverage: 75% (below 80% threshold)
• Flag: "network_degraded_session"
• Action: Offer session retry at no cost, debrief marked as "practice only"

Example 3: Proctor False Positive Correction

Setup: User wears glasses, has second monitor for work (turned off during session).

Session Flow:

05:00 MediaPipe detects gaze deviation (actually lens reflection)
05:01 Proctor flags "off-screen gaze"
05:02 User continues normally
08:00 Another false flag (glasses adjustment)
08:01 System detects: flag_rate=2 in 8 minutes = high
08:05 Auto-calibration triggered: "We're detecting gaze issues. 
      Let's recalibrate for your setup. Look at the center of 
      your screen for 10 seconds."
08:15 New baseline established for this user
      Subsequent flags drop to normal levels

Quality Gates

Session completion requires all checks to pass:

• [ ] Voice stream integrity >80% (no major dropouts)
• [ ] Transcript coverage >80% (enough content for evaluation)
• [ ] Session duration within 10% of target (not cut short)
• [ ] At least 3 meaningful question-response exchanges
• [ ] Emotion data captured for >70% of user responses
• [ ] Whiteboard snapshots captured every 30s during active drawing (design rounds)
• [ ] Proctor calibration successful if first-time user
• [ ] No critical API failures (can fallback gracefully)
• [ ] User completed session normally (didn't force-quit)
• [ ] Debrief generation successful with specific feedback

Not-For Boundaries

This skill should NOT be used for:

• Individual round coaching: For practicing specific round types, use the round-specific skills (senior-coding-interview, ml-system-design-interview, etc.)
• Interview strategy/timeline: For prep planning and round sequencing, use interview-loop-strategist instead
• Resume/portfolio work: For CV optimization, use cv-creator instead
• Salary negotiation: Different skill set, not interview simulation
• Conference presentations: Different audience and evaluation criteria than interviews
• Academic exam proctoring: This is for self-practice, not adversarial examination
• Production hiring platform: This is for candidate practice, not company interview infrastructure

Delegate to other skills when:

• User asks about specific technical topics → Use the relevant round-specific skill
• User wants to plan their overall prep strategy → Use interview-loop-strategist
• User needs help with career positioning → Use career-biographer or cv-creator