Blabber Coach

Full-voice mock interview coach. Upload a resume, choose a domain and difficulty, speak your answers BlabberCoach listens, pushes back, then tells you exactly what to fix.

Comment

Building BlabberCoach: A Voice-First AI Interview Coach

Inspiration

BlabberCoach started from a pattern I kept noticing while helping friends prepare for interviews.

Many candidates are technically strong but still struggle during live interviews. The issue usually isn’t knowledge — it’s delivery under pressure: pacing, filler words, confidence, and clarity.

Most interview preparation tools are text-first. They help candidates write answers, but they don’t train the part that often determines interview outcomes: spoken communication.

So I decided to build BlabberCoach, a voice-first interview simulator that lets candidates practice interviews the way they actually happen — through real-time conversation.

System Design

BlabberCoach is built as a multi-agent AI pipeline that simulates a realistic technical interview.

The system consists of several coordinated stages.

Resume Parsing and Profile Extraction

The system begins by analyzing the candidate’s resume to extract relevant signals such as:

  • skills
  • technologies
  • experience level
  • target roles

This enables the interview to be personalized to the candidate's background rather than relying on generic questions.

Intelligent Question Generation

Questions are generated dynamically based on:

  • role or domain
  • candidate experience level
  • selected difficulty

This allows the system to produce context-aware technical questions and adaptive follow-ups.

Live Voice Interview

The core feature is a real-time bidirectional voice interview.

During the interview the system:

  • transcribes the candidate’s speech
  • maintains conversation context
  • generates follow-up questions
  • adapts the conversation flow

The goal is to create an experience closer to a real technical interview rather than a chatbot interaction.

Post-Interview Evaluation

After the session ends, BlabberCoach generates a structured feedback report including:

  • quote-linked feedback
  • communication analysis
  • answer rewrites
  • actionable improvement suggestions

This helps candidates clearly understand how to improve before their next interview.

What I Learned

One of the biggest lessons from building BlabberCoach was that AI reliability matters as much as model capability.

To make the system reliable in production-like workflows, several safeguards were necessary:

  • strict schema validation for model outputs
  • prompt contracts to maintain predictable responses
  • graceful fallbacks when validation fails

Without these guardrails, even powerful models can create fragile systems.

Another important realization was how much UX sequencing influences perceived performance.

For example:

  • performing interview setup while precomputing question plans
  • providing progressive analysis feedback
  • visualizing results with clear scoring rubrics

These improvements significantly increased user trust and responsiveness.

Challenges

Building a real-time AI interview system introduced several technical challenges.

Real-Time Voice Orchestration

Maintaining natural conversation required coordinating:

  • speech recognition
  • language model responses
  • voice synthesis

while keeping latency low enough for fluid conversational turn-taking.

Infrastructure Constraints

The system needed to work within serverless environments and WebSocket-based communication, which required:

  • managing session state
  • handling real-time message streams
  • implementing environment-based fallbacks

Evaluation Quality

Another difficult challenge was converting qualitative speech signals into useful metrics without introducing noise.

The scoring system had to balance measurable signals with contextual understanding.

Some signals are intentionally explicit.

Scoring Signals

Filler Word Rate

$$ \text{Filler Rate} = \frac{\text{filler_count}}{\text{total_words}} \times 100 $$

This measures the percentage of filler words such as um, uh, or like.

Words Per Minute

$$ \text{WPM} = \frac{\text{word_count}}{\text{duration_seconds}/60} $$

This evaluates speaking pace to determine whether a candidate is speaking too quickly or too slowly.

Overall Interview Score

$$ \text{Overall Score} = 0.40C + 0.25Q + 0.20V + 0.15D $$

Where:

  • (C) = Content quality
  • (Q) = Communication clarity
  • (V) = Vocabulary and articulation
  • (D) = Delivery and confidence

These signals help make scoring transparent and explainable rather than a black-box evaluation.

Final Thoughts

BlabberCoach was built to address a simple but overlooked problem: technical preparation rarely trains verbal performance.

By combining voice interaction, AI evaluation, and explainable feedback, the goal is to make interview preparation closer to the real experience candidates face.

Building the project reinforced an important lesson: effective AI products depend not only on model capability, but on reliable system design around the model.

Built With

Share this project:

Updates