I Built an AI Interview Coach with Gemma 4 — Zero Backend, 100% Free

This is a submission for the Gemma 4 Challenge: Build with Gemma 4.

What I Built

Interview Coach — a free, open-source AI interview practice tool powered by Google Gemma 4 (31B Dense). It runs entirely in the browser — no backend, no server, no accounts. Just Gemma 4's reasoning and your ambition.

It conducts realistic mock interviews across 6 modes, evaluates your answers in real-time with detailed feedback, gives mid-session scorecards, and generates comprehensive performance reports with personalized study plans.

The world's first free and open-source AI interview coach.

Demo

🔗 Live Demo: hajirufai.github.io/gemma4-interview-coach (bring your own free API key from Google AI Studio)

💻 GitHub Repo: github.com/hajirufai/gemma4-interview-coach

📄 License: MIT — fork it, improve it, ship it.

The Problem

91% of candidates who fail online assessments never practiced under realistic conditions. Interview prep tools exist, but they're:

• Expensive — $30–50/month for premium platforms
• Generic — same questions regardless of your role or level
• Passive — read sample answers instead of actually practicing

What if you had a personal coach that adapts to YOUR experience level, gives feedback on YOUR specific answers, and costs $0 forever?

That's what Gemma 4 makes possible.

6 Practice Modes

Each mode has a carefully crafted system prompt that shapes Gemma 4's behavior — asking follow-ups, evaluating with mode-specific criteria, and calibrating difficulty across entry/mid/senior/lead levels.

Key Features

🎙️ Voice Input & Output

Speak your answers naturally using Web Speech API — just like a real interview. Gemma 4's responses are read aloud with text-to-speech for a fully conversational experience.

📊 Real-Time Scoring

Hit "Score Me" at any point for a mid-session evaluation across 5 dimensions. The end-of-session report includes a detailed performance breakdown and a personalized 7-day study plan.

🖼️ Image Upload

Upload screenshots of coding challenges, whiteboard diagrams, or error messages — Gemma 4's multimodal capabilities analyze them in the context of your interview session.

🔄 Multi-Provider Support

Works with Google AI Studio, OpenRouter, NVIDIA NIM, and HuggingFace — all using Gemma 4. If one provider is overloaded, switch seamlessly.

✅ API Key Validation

One-click key test before starting — instant feedback on whether your setup is working.

📋 Session History

LocalStorage-backed session tracking so you can see your progress over time — modes practiced, questions completed, and token usage.

🌙 Dark Mode

Full dark/light theme toggle — easy on the eyes during late-night interview prep.

How I Used Gemma 4

Model Choice: 31B Dense — And Why It Matters

I deliberately chose the 31B Dense model over the lighter MoE variants. Here's why:

Interview coaching demands the highest reasoning quality available. When Gemma 4 evaluates a candidate's answer, it needs to simultaneously:

• Parse whether the answer follows frameworks like STAR or MECE
• Identify specific gaps ("You mentioned leading the migration, but what was the quantifiable impact?")
• Calibrate difficulty for the next question based on performance so far
• Generate feedback that's encouraging but honest

The 31B Dense model activates all 31 billion parameters for every token, producing noticeably more nuanced and accurate evaluations than the smaller variants. For a coaching tool where feedback quality IS the product, this was non-negotiable.

128K Context Window → True Multi-Turn Coaching

This is where Gemma 4 really shines. Interview practice isn't a one-shot Q&A — it's a 15–20 turn conversation where the coach needs to:

• Remember your answer to Q1 when evaluating Q8
• Notice patterns ("You keep avoiding specifics — let me push harder")
• Generate a final report that references the entire session

System prompt:     ~800 tokens
Per Q&A round:     ~500 tokens (question + answer + feedback)
15 rounds:         ~7,500 tokens
Final report:      ~2,000 tokens
Total:             ~10,300 tokens ← comfortably within 128K

No chunking, no summarization, no lost context. Every answer is remembered and referenced.

Native Chain-of-Thought → Superior Evaluations

Gemma 4's built-in thinking tokens (thought: true) are game-changing for evaluation tasks. Before responding, the model reasons internally:

{
  "text": "The user mentioned leading a team of 5...\n- STAR compliance? Situation ✓, Task ✓, Action partial, Result missing\n- Specificity? Medium — needs quantified metrics\n- Overall assessment: push for more concrete outcomes",
  "thought": true
},
{
  "text": "Great foundation! You clearly described the situation and your role. To make this a knockout answer, add the specific outcome..."
}

This internal reasoning produces dramatically better feedback than models that generate evaluations in a single pass.

Open & Free → Accessible to Everyone

This was the whole point. Gemma 4 runs on:

• Google AI Studio free tier — no credit card required
• OpenRouter — free tier available
• Locally via Ollama — on a decent laptop
• NVIDIA NIM / HuggingFace — for developers who prefer those platforms

Zero cost, zero gatekeeping, zero excuses not to practice.

Architecture: Why Zero Backend?

Browser ──(HTTPS)──> Gemma 4 API
   │                    │
   │              31B Dense Model
   │                    │
   └────────────────────┘
     No server. No proxy. No database.

The entire app is a single HTML file. No React build, no Node server, no database. Here's why:

1. Privacy — Your API key and interview responses never touch a third-party server
2. Cost — $0 hosting via GitHub Pages
3. Speed — No proxy round-trip. Browser → Gemma 4 → Browser
4. Simplicity — git clone && open index.html is the full setup

The tradeoff: Users need their own API key. I chose this intentionally — it keeps the tool free forever and teaches users about AI APIs in the process.

Tech Stack

What I Learned

1. Gemma 4's thinking tokens are game-changing for evaluation tasks. The model genuinely considers multiple aspects before responding — producing feedback that feels like a real interviewer's assessment, not a generic LLM response.

2. 128K context is overkill for most apps — but perfect for coaching. The ability to reference earlier answers creates a coherent experience that shorter-context models can't match. The final session report can cite specific moments from Q1 through Q15.

3. Zero-backend AI apps are viable and powerful. Browser → API → Browser eliminates 90% of infrastructure complexity. The main cost is that users bring their own key — but for free tools, that's a feature, not a bug.

4. Multi-provider resilience matters. Google AI Studio's free tier occasionally returns 500/503 errors under load. Having OpenRouter, NVIDIA NIM, and HuggingFace as fallbacks — all serving Gemma 4 — keeps the tool reliable.

Try It

🔗 Live Demo: hajirufai.github.io/gemma4-interview-coach
💻 Source Code: github.com/hajirufai/gemma4-interview-coach
⭐ Star the repo if you find it useful!

Built by Haji Rufai — creator of Interview Buddy, a free AI-powered interview preparation platform.