🎙️ AI Voice Interviewer

Practice job interviews out loud with an AI interviewer, then get scored, written feedback in seconds.

Live demo

What it does

• Builds a custom interview by voice. Tell the agent the role, seniority, tech stack and how technical you want it. Gemini writes the questions.
• Runs a real-time spoken interview. Low-latency speech-to-text, an LLM interviewer that asks follow-ups, and a natural voice, all over WebRTC through Vapi.
• Scores you on five categories. Communication, technical knowledge, problem solving, culture fit and confidence, each 0 to 100 with comments, plus strengths, areas to improve and an overall verdict.
• Keeps your history. Every interview and its feedback is saved, so you can retake it and track progress. Interviews other users generate are available to practice too.

Architecture

The app is a single Next.js App Router project with no separate backend. The browser streams audio to Vapi, which handles speech-to-text, the interviewer LLM and text-to-speech. When the "generate" voice workflow has collected your preferences, Vapi calls a webhook in this app that asks Gemini for questions and stores them in Firestore. When an interview ends, a server action sends the transcript to Gemini with a typed zod schema and saves the structured feedback.

flowchart LR
    B[Browser<br/>Next.js + Vapi Web SDK] <-- voice over WebRTC --> V[(Vapi<br/>STT · LLM · TTS)]
    V -- workflow webhook --> W["/api/vapi/generate"]
    B -- transcript --> SA[Server action<br/>createFeedback]
    W --> C[lib/interview<br/>prompts · parsing · schemas]
    SA --> C
    C --> G[(Gemini 2.0 Flash)]
    W --> F[(Firestore)]
    SA --> F

Sequence diagrams for both voice flows, the auth model and the data model are in docs/ARCHITECTURE.md.

Tech stack

Layer	Tools
Frontend	Next.js 15 (App Router), React 19, TypeScript, Tailwind CSS 4, shadcn/ui
Voice	Vapi (Deepgram Nova-2 STT, GPT-4 interviewer, ElevenLabs TTS)
LLM	Google Gemini 2.0 Flash via the Vercel AI SDK (generateText, generateObject)
Auth and data	Firebase Auth with httpOnly session cookies, Firestore
Validation	zod for request bodies and structured LLM output
Quality	Vitest, ESLint, GitHub Actions

Run locally

Prerequisites: Node.js 20.19+ or 22.12+, a Firebase project (Auth with email/password, Firestore), a Vapi account and a Gemini API key.

git clone https://github.com/obro79/mock_interview_ai.git
cd mock_interview_ai
npm install
cp .env.example .env.local   # then fill in the values
npm run dev                  # http://localhost:3000

Variable	Where it comes from
NEXT_PUBLIC_FIREBASE_*	Firebase console → Project settings → Your apps
FIREBASE_PROJECT_ID, FIREBASE_CLIENT_EMAIL, FIREBASE_PRIVATE_KEY	Firebase console → Service accounts → Generate new private key
GOOGLE_GENERATIVE_AI_API_KEY	Google AI Studio
NEXT_PUBLIC_VAPI_WEB_TOKEN	Vapi dashboard → API keys (public key)
NEXT_PUBLIC_VAPI_WORKFLOW_ID	Vapi dashboard → your "generate interview" workflow

The Vapi workflow's API request node should POST to https://<your-host>/api/vapi/generate with role, level, techstack, type, amount and userid.

Tests

npm test            # Vitest unit tests
npm run typecheck   # tsc --noEmit
npm run lint        # ESLint (next/core-web-vitals)

The suite covers the core logic in lib/interview and lib/tech: prompt construction, parsing messy LLM output into questions, request validation and the feedback schema. CI runs all three checks plus a production build on every push and pull request.

Credits

Built by Owen. The UI and base app structure started from the JavaScript Mastery AI interview platform walkthrough. Extended design notes are on Notion.

Licensed under MIT.