DEV Community

Arpita Singh
Arpita Singh

Posted on Edited on

I Built Rehearse to Sell A Voice-First, Local-First AI Interview Coach

Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle Submission 💸

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

I built Rehearse, a voice-first and local-first AI interviewer.

You know how you read a list of "Top 50 Interview Questions," think "I could answer that," and then completely freeze when you actually have to speak out loud? That is the exact problem this solves.

ChatGPT doesn't help here. If you ask it to practice with you, it gives you a giant essay, says "Great job!", and never pushes back. A real interview is a conversation. Rehearse actually talks to you (using voice), cuts you off if you ramble, and forces you to answer the hard follow-up questions. At the end, it grades you out of 10 and tells you how a better candidate would have answered.

I didn't want this to just be for software engineers. You can type "Marketing," "Nursing," or "Sales," and paste a real job description (I tested it with a Deloitte JD), and the AI changes its tone and questions to match that field. And if you pick "Stress Test" mode, it acts like a genuinely hostile interviewer.

Demo

🔗 Live Demo: https://rehearse-g3q4.onrender.com

Demo Image 01

Demo Image 01

Demo Image 02

Demo Image 02

Demo Image 03

Demo Image 03

Demo Image 04

Demo Image 04

Code

📊 GitHub Repository: https://github.com/arpitasi1gh/hacktoberfest/tree/main/rehearse

How I Built It

The stack is FastAPI, HTMX, and Gemma. I used the browser's built-in microphone and text-to-speech so it works instantly without saving audio files to a server.

The hardest part wasn't the code, it was stopping the AI from being too nice. I had to write a very strict set of rules for the prompt. It is explicitly banned from saying things like "Great answer" or "I understand." If you say "we built a system," it is instructed to interrupt you and ask, "What did you specifically do?"

Then I ran into two hilarious bugs.

First, Gemma on Google's API is a "thinking" model. Instead of just asking the question, it would write out its entire internal thought process: "Wait, the prompt says no meta-text. Let's try asking this..." Because my app reads the text out loud, it was literally speaking the AI's internal self-doubt to the user.

Second, when I pasted a massive job description into the context field, the local model got confused and started wrapping its JSON output in markdown backticks (

json {...}

). My UI just dumped the raw JSON onto the screen. I had to write a bulletproof Python parser that scans the AI's output bottom-up, ignores the monologue, strips the markdown, and grabs only the final {"question": "..."} object.

Why Does Open Innovation Matter?

Interviewing is vulnerable. You are anxious, alone, and talking to a screen. If this used a closed API, it would mean sending your awkward voice recordings to a giant tech company's server.

Because it uses open-source AI, this app can run 100% locally. If you run it with Ollama, nothing leaves your laptop.

To make the app demoable as a live link (since free hosting can't run a 4GB model), I built a fallback: it tries the local Ollama server first, and if you don't have it running, it uses the Google API. This local-first setup is only possible because open models are finally small enough to run on normal laptops.

Prize Categories

  • Best Use of Gemma — Gemma 3 4B via local Ollama, with a Google AI Studio Gemma fallback.
  • Best Use of Render — FastAPI app deployed on Render using the cloud inference fallback.

Top comments (0)