Practicing Interviews Without Uploading Your Career

Job interviews are a skill. Like any skill, practice helps. But how do you practice answering "Tell me about a time you handled conflict" without:

1. Bothering friends to roleplay
2. Paying $200/hour for a career coach
3. Uploading your resume to yet another service

I built HotSeat AI to solve this problem entirely on my own hardware.

The Privacy Stakes

Interview preparation involves sensitive information:

• Your complete work history
• Companies you're targeting
• Salary expectations
• Recorded practice sessions with your face/voice
• Honest assessments of your weaknesses

Cloud-based interview prep tools collect all of this. Some sell it for "recruitment insights." Others get breached. All of them have your data forever.

Running locally means: your career stays your business.

The Components

Building a full interview coach for local deployment was an ambitious project that evolved multiple times as the requirements for various AI capabilities changed:

Voice Conversation

• Speech-to-Text: whisper.cpp transcribes your spoken answers
• Text-to-Speech: Piper TTS gives the AI a voice
• LLM: Ollama (Gemma 3 or Qwen 2.5) generates questions and feedback

Body Language

MediaPipe runs in the browser, analyzing your webcam feed:

• Posture (are you slouching?)
• Eye contact (looking at camera vs. away)
• Nervous movements (excessive fidgeting)

Resume Analysis

The same LLM that conducts interviews can:

• Score ATS compatibility
• Find keyword gaps vs. job descriptions
• Suggest professional improvements

How a Session Works

1. Select Mode

   • Quick Start: Random questions
   • Targeted: Paste a job description
   • Custom: Choose question categories

2. Choose Persona

   • 20+ AI interviewers: friendly, tough, technical
   • Different industries and seniority levels

3. Practice

   • AI asks a question via voice
   • You respond naturally
   • Body language feedback
   • AI follows up based on your answer if necessary

4. Review

   • Phase-by-phase breakdown
   • Scores: clarity, confidence, technical accuracy, STAR method
   • Specific suggestions for improvement

Technical Challenges

Latency Matters

Interviews are conversational. If the AI takes 10 seconds to respond, the practice feels unnatural. Optimizations:

• Streaming transcription (starts processing before you finish speaking)
• Aggressive model quantization (smaller = faster)
• Response streaming (starts speaking before generation completes)

With decent hardware, the loop feels close to real-time.

Multi-Modal Coordination

The app juggles several simultaneous streams:

• Microphone → whisper.cpp → text
• Text → LLM → response
• Response → TTS → audio
• Camera → MediaPipe → body language metrics

Each has different latency characteristics. Synchronizing them required careful async programming.

Hardware Requirements

CPU-only works but responses will be slower. A mid-range GPU or Apple Silicon makes conversations feel natural.

The Feedback Loop

After each session, the app provides detailed analytics:

Per-Question Scores

• How clearly did you communicate?
• Did you use specific examples?
• Was your answer structured (STAR method)?

Behavioral Competencies

• Leadership examples
• Conflict resolution
• Technical problem-solving

Progress Over Time

• Trend charts across sessions
• Identify persistent weak areas
• Track improvement

What I Learned

Local AI is Conversational

I was skeptical that local models could hold realistic interviews. They can. Gemma 3 4B asks probing follow-ups, catches contradictions, and provides specific feedback.

Body Language Analysis Works

MediaPipe is surprisingly accurate at detecting posture and eye contact. The post-practice report showing "low eye contact" made me immediately aware of habits I hadn't noticed, changing how I approached my next session.

Privacy Enables Honesty

Knowing no one else would see my practice sessions made me more willing to be honest about weaknesses. I practiced answering difficult questions I'd never record on a cloud service.

Results

The app delivers on its core promise:

• Realistic interview practice with AI-generated follow-up questions
• Immediate feedback on answer structure and body language
• Progress tracking across sessions to identify improvement areas
• Complete privacy with zero data leaving your device

No career coach fees. No data uploaded anywhere. Just practice, feedback, improvement.

Source Code

HotSeat AI is open source: github.com/fullstackgrower/fsg-hotseat-ai