How We Built Aura: A Real-Time AI Interview Coach with Gemini Live, ADK, and Vertex AI

This article was created for the purposes of entering the Gemini Live Agent Challenge. #GeminiLiveAgentChallenge

🎙 Try Aura Live

🌐 Live Demo: https://aura-sde-interview-agent-ivhauk7c7a-uc.a.run.app
💻 GitHub Repo: https://github.com/srthorat/aura-sde-interview-agent

The Biggest Barrier Isn’t Skill. It’s Interview Fear.

Many talented candidates get shortlisted for top tech companies, but never reach their full potential in interviews.

Not because they lack technical ability.

Because they lack realistic interview practice.

Mock interviews are expensive, difficult to schedule, and often inaccessible for students outside elite networks or major cities.

Most preparation resources are text-based.

People read. Few truly practice.

We wanted to change that.

So we built Aura — a real-time AI interview coach powered by Google AI models and Google Cloud.

No text box. Just an interview.

⚡ What is Aura?

Aura is a voice-first interview simulator that lets candidates practice technical interviews naturally through conversation.

Candidates can:

• Practice coding rounds
• Practice system design rounds
• Practice behavioral interviews
• Receive spoken feedback instantly
• Get live grading updates
• Return later and continue where they left off

Aura behaves like a real interviewer — not a chatbot.

🧠 Why Voice-First Matters

Interviewing is a spoken skill.

You need to think aloud, structure answers, handle pressure, communicate tradeoffs, and respond in real time.

Typing into a textbox cannot simulate that.

Aura uses live voice conversation so candidates can build confidence under realistic conditions.

🏗 System Architecture

Core Stack

Frontend

• React + Vite
• LiveKit JS SDK

Transport

• LiveKit WebRTC for low-latency audio streaming

Backend

• FastAPI (Python 3.11)
• Containerized on Cloud Run

AI Layer

• Gemini Live (gemini-live-2.5-flash-native-audio)
• Google ADK agent orchestration + tools

Memory Layer

• Vertex AI session persistence

Infra

• Cloud Build
• Artifact Registry
• Secret Manager

🔊 Why We Chose Gemini Live

We wanted natural, interruptible conversation.
Gemini Live gave us:

• Real-time bidirectional voice streaming
• Fast spoken responses
• Interruption / barge-in handling
• Tool calling during conversation
• Native voice experience without stitching multiple systems together This made Aura feel like a real interviewer.

🛠 How Google ADK Helped

We used Google ADK to structure interview workflows with reusable tools such as:

• evaluate_candidate_answer
• submit_rubric_grade
• get_round_scorecard
• get_session_summary
• end_conversation This let us combine model intelligence with deterministic scoring logic.

💾 Persistent Candidate Memory with Vertex AI

Aura supports returning candidates.
It can remember:

• Previous interview rounds
• Prior scores
• Weakness areas
• Suggested next practice topics
• Improvement over time This transforms Aura from a one-time mock interviewer into a long-term coaching platform.

☁️ Why Cloud Run

We wanted fast iteration and production-ready deployment.
Cloud Run gave us:

• Containerized deployments
• Autoscaling
• Low operational overhead
• Easy CI/CD integration
• Reliable public hosting for demos Perfect for hackathon speed and startup scalability.

🚀 What We Learned

1. Voice UX matters more than UI polish
2. Latency is everything in real-time AI
3. Persistent memory creates user stickiness
4. Interview confidence is trainable
5. Google AI stack dramatically accelerates development

🌍 Bigger Vision

Aura starts with interview coaching.
The same architecture can power:

• Sales roleplay coaching
• HR screening simulators
• Language learning tutors
• Customer support training
• Certification prep agents

Closing

We believe the future of learning and preparation will be conversational, adaptive, and always available.
Aura is one step toward that future.

Aura
No text box. Just an interview.

• #GeminiLiveAgentChallenge #GoogleCloud #Gemini #VertexAI #AI #VoiceAI #Hackathon