AeroPrep

Spoken mock interviews with an AI interviewer that has read your resume.

You upload a resume (or describe the role), talk to "Alex" for about fifteen minutes, optionally solve one coding problem, and get a scored report with a note on every answer. The first three interviews on an account are free.

Live site: https://ai-interview-coach-eight-mu.vercel.app

What it does

• Interviews from your own resume. Questions are about the projects and tools you listed, pitched at the experience level you choose.
• A real conversation. Alex follows an interview arc (project walk-through, deep dive, fundamentals, a scenario, a behavioural question, your questions), asks follow-ups when an answer is thin, and replies sentence by sentence as the answer is generated.
• Voice or keyboard. Speak (Chrome and Edge) or type. A spoken answer is sent after a pause whose length you choose.
• Your resume stays yours. The PDF is read on the server and only its text is kept; the file is never stored.
• Nothing lost on a refresh. The transcript and the clock live on the server. Close the tab mid-interview and you resume at the same question.
• A coding round. One challenge per interview, reviewed by the model. It is not executed; see "Ideas" below.
• A report that quotes you. Scores, strengths, weaknesses, a hire or no-hire call, and feedback on each answer.
• Attention nudges. With the camera on, the page notices if you leave the frame, look away for several seconds, or switch tabs. This runs entirely in the browser; no video leaves the device.

How it is built

Part	Stack	Where
Frontend	Next.js 16 (App Router), React 19, Tailwind 4	frontend/, deployed on Vercel
Backend	Node 22, Express 5, Socket.IO, Prisma	backend/, deployed on Render
Database	PostgreSQL	Neon
Model	Any provider, picked from the API key	backend/services/llm/

The browser holds only a login token and the id of the current interview. The server owns the transcript, the clock and every limit, and makes every model call itself. ARCHITECTURE.md walks through a session step by step.

Run it locally

You need Node 20 or newer and a PostgreSQL database.

Backend

cd backend
npm install
cp .env.example .env      # then fill in DATABASE_URL, JWT_SECRET and AI_API_KEY
npm run migrate           # creates the tables
npm run dev               # http://localhost:5001

Frontend

cd frontend
npm install
echo "NEXT_PUBLIC_API_URL=http://localhost:5001" > .env.local
npm run dev               # http://localhost:3000

Configuration

Backend (backend/.env):

Variable	Needed	What it does
DATABASE_URL	yes	PostgreSQL connection string
JWT_SECRET	yes	Signs logins. Without it a temporary secret is used and everyone is signed out on restart
AI_API_KEY	yes	Key for the model provider. The provider is picked from the key: Anthropic, OpenAI, OpenRouter, Groq or Gemini
AI_MODEL	yes, except Anthropic	Model id. Defaults to claude-opus-5-5 for Anthropic
AI_MODEL_FAST	no	A quicker, cheaper model used only for the live interview turns
AI_EFFORT	no	Anthropic only: low, medium or high. low is faster and cheaper
AI_PROVIDER, AI_BASE_URL	no	Force a provider, or point at any OpenAI-compatible endpoint (custom, ollama)
FREE_INTERVIEW_LIMIT	no	Interviews per account. Default 3; User.interviewLimit overrides it per user
DAILY_INTERVIEW_LIMIT	no	Interviews that may start per day across all users. Default 100. A ceiling on spend
INTERVIEW_MINUTES, INTERVIEW_MAX_ANSWERS, CODING_MAX_RUNS	no	Length of an interview (15), answers in it (12), code evaluations (5)
INTERVIEW_MAX_FAILED_CALLS	no	Model calls per interview that may fail after possibly being paid for (a timeout, an unusable reply) before the interview stops making calls. Default 6
VOICE_API_KEY, VOICE_REGION, VOICE_NAME	no	A speech service (ElevenLabs, Deepgram, Azure or an OpenAI-style service; up to three, tried in order, with VOICE2_ and VOICE3_ settings for the fallbacks) for a natural interviewer voice. Without a key the browser's built-in voice is used. VOICE_DAILY_CHARS (default 60000) caps daily use
SMTP_USER, SMTP_PASS	no	Gmail address and app password; in-app feedback is emailed to SUPPORT_EMAIL
SUPPORT_EMAIL	no	Address shown to users and that receives feedback
CORS_ORIGINS	no	Sites allowed to call the API besides the built-in ones, comma separated. Set this if you host the frontend anywhere other than the default address
MONITOR_TOKEN	no	When set, /monitor and /metrics need ?token=

Frontend (frontend/.env.local): NEXT_PUBLIC_API_URL, and optionally NEXT_PUBLIC_SITE_URL (the site's own address, used in link previews). Anything in a NEXT_PUBLIC_ variable is sent to the browser, so never put a secret in one.

Cost

Every model call is made by the server, against caps: one call per answer, at most 12 answers, one coding challenge with 5 evaluations, one report. Calls that fail are bounded too: a reply that was paid for is kept and counted even when it is empty or breaks off, and an interview stops calling the model after a handful of failures that may have been charged. Token use is recorded per interview and shown on /monitor. On Claude Opus 5.5 with AI_EFFORT=low, a full interview with its report costs roughly ten US cents; setting AI_MODEL_FAST to a smaller model cuts that substantially.

Tests

cd backend
TEST_DATABASE_URL=postgresql://postgres:postgres@localhost:5432/aeroprep_test npm test

The suite starts the real server against that database and a stand-in for the model, and drives it over HTTP and Socket.IO. It empties the database it is given, so point it at a throwaway one.

Deploying

• Backend (Render): root directory backend, build npm install, start npm start. After pulling a change that adds a migration, run npm run migrate against the production database before the new code starts.
• Frontend (Vercel): root directory frontend. Set NEXT_PUBLIC_API_URL to the backend's URL.
• When the API between them changes, let the backend finish deploying before the frontend goes out: the new pages call routes the old server does not have.

Ideas

Things we would like to build, roughly in order of value:

1. Run code for real in a sandbox instead of asking the model to judge it.
2. A more natural interviewer voice than the browser's built-in one.
3. Email verification and password reset.
4. System design and behavioural-only interviews, and company-style question sets.
5. Progress over time: scores across interviews and recurring weaknesses.

If you want to help, or have a better idea, write to [email protected] or open an issue.