● InterviewCoach AI · open source · MIT

Walk into every interview already rehearsed.

A private, open-source rehearsal studio: real-time voice interviews with an AI that knows your resume, feedback that refuses to lie, coding practice, and a system-design arena — running locally in your browser.

🎙 Voice interviews 📄 Resume-aware 🔒 Local-first privacy ⚖ MIT open source

View on GitHub Quick start →

localhost:1234/dashboard

InterviewCoach AI Performance Hub dashboard

The Performance Hub — trajectory, skill radar, pressure index, across every session you've run. real screenshotReal screenshotCaptured from the actual InterviewCoach AI interface, shipped in the repo's assets/. Shown here with the app's built-in mock data.

practice modes — role, seniority, difficulty, duration, company track, campus, blind, demo

built-in company tracks, from campus placements to Google and Amazon loops

practice pillars — voice interviews, coding arena, system-design canvas

cloud sync. Your interviews, reports and resume stay in your browser

The problem

Interviews are won before you walk in.

And almost nobody gets to rehearse under anything close to real conditions. That's the gap InterviewCoach AI exists to close.

🎯

High-stakes, low-rep

You get a handful of real interviews a year, and each one bends months of your trajectory. The pressure is real; the chance to practice under it is nearly impossible to arrange.

💸

Friends are awkward, coaches are expensive

Mock interviews with friends are kind and forgettable. Professional coaches charge real money per session — and they're booked out when you need them most.

📋

Question banks don't know you

Generic lists ask generic questions. Nobody asks about the migration on your resume, the outage you debugged — until the real interviewer does, live, while you're sweating.

Hero feature

Talk. Get interrupted. Recover. Like the real thing.

Real-time voice interviews over the Gemini Live API — not a chatbot with a microphone taped on. The interviewer speaks, listens, and pushes back while you're still talking.

⚡
Natural interruption
Ramble, and the interviewer cuts in — the way a real one would. Conversation, not turn-taking.
📝
Live transcription
Every word captured as it's spoken, so nothing you said well gets lost to nerves.
⏸
Pause, resume, reconnect
Life interrupts. Pause mid-answer, resume cleanly — and if the connection drops, the session reconnects and recovers instead of dying.
🎧
Hardware preflight
Before you start: record your mic, play it back, verify the signal. No discovering your audio was dead five minutes into a mock.

Voice pipeline from docsFrom the docsCapabilities stated in the project's README and implemented over the Gemini Live API.

localhost:1234/interview · live session

Live voice interview session with interviewer Sarah

A live session: the voice orb, session timer, live transcript panel, and controls. real screenshotReal screenshotCaptured from the actual InterviewCoach AI interface, shipped in the repo's assets/.

The voice orb, stylized. illustrativeIllustrativeA stylized depiction of the live-audio visual, not a product capture.

Resume-aware

It read your resume. All of it.

localhost:1234/dashboard · resume intelligence

Resume Intelligence view with AI summary and topic confidence

Resume Intelligence: an AI summary of your strengths, topic-level confidence, and what to practice next. real screenshotReal screenshotCaptured from the actual interface. The resume shown is the app's sample data.

Drop in a PDF and the interviewer stops asking "tell me about yourself" and starts asking about your migration, your outage, your trade-offs — the questions the real interviewer will actually ask.

📄
Parsed locally, in your browser
Resumes are extracted with a local PDF.js worker — 5 MB and 20 pages max. Your file is only ever sent to Gemini as interview context, never stored anywhere.
🎯
Questions about your experience
The session is grounded in what you've actually built. Defend your own decisions, explain your own architecture.
📊
Topic confidence, honestly shown
The dashboard shows where your answers hold up and where they wobble — per topic, with suggested practice for each gap.

Resume pipeline from docsFrom the docsLimits and behavior stated in the project's README.

Feedback

Feedback that refuses to lie.

Every report is validated before it's displayed or saved. Missing or malformed scores are not fabricated — if the evidence isn't there, the report says so.

🔍
Evidence-validated reports
Scores trace back to what you actually said. No confident-sounding numbers invented to fill a template.
⭐
STAR review
Your behavioral answers are broken down Situation → Task → Action → Result, so you can see exactly where the story lost its shape.
🗺
Study plans that target you
Weak on system-design trade-offs? The plan says so, and generates focused sessions to fix exactly that.
📈
Progress history
Session-over-session tracking — pressure resilience, skill evolution, interview frequency — so you know whether you're actually improving.

localhost:1234/dashboard · strengths & improvements

Top strengths, priority improvements and interview frequency

Strengths, priority improvements with one-click focused sessions, and your practice cadence. real screenshotReal screenshotCaptured from the actual InterviewCoach AI interface, shipped in the repo's assets/.

Practice arena

A coding gym with a spotter.

JavaScript challenges run in a disposable, time-bounded browser Worker — three seconds, then it's gone. Built for interview practice, with coaching layered on top.

⚙️

Safe-by-design runner

Your functions execute against JSON-compatible tests inside a throwaway Worker with a 3-second timeout. It's a practice gym, not a sandbox for untrusted code — Python, Java and C++ stay out by design.

🩺

Skill diagnostics

An optional diagnostic maps what you can actually do, topic by topic, instead of guessing from a résumé line.

📚

Ranked resources

Learning resources ranked for your gaps, problem recommendations matched to your level, and growth tracking that shows the curve moving.

Arena behavior from docsFrom the docsRunner scope and behavior stated in the project's README.

System design

Whiteboard rounds, minus the whiteboard anxiety.

A full design canvas for the round most candidates fear most — with coaching and scoring built in, so your boxes-and-arrows actually get critiqued.

🔍

Navigate freely

Anchored wheel and button zoom, panning, fit and reset view, plus a minimap so you never lose the big picture mid-design.

🧲

Draw neatly, fast

Snap grid, auto-layout, starter architectures, duplicate, and full undo/redo. Labels and notes keep the reasoning attached to the diagram.

💾

Take it with you

Canvases persist locally and export or import as JSON; diagrams export as SVG. Coaching and scoring tell you what to fix before the real loop.

Canvas capabilities from docsFrom the docsFeature list stated in the project's README.

Every way to practice

Eight ways to walk in nervous.

Dial in exactly the interview you're afraid of — or let the app surprise you.

💼

Role

Target the actual position — Software Engineer today, whatever's next tomorrow.

📶

Seniority

Junior, Mid-Level, Senior, Executive. The questions change because the bar does.

🎚

Difficulty

Easy, Medium, Hard. Warm up or get humbled — your call.

⏱

Duration

5, 10, 15, 20 or 30 minutes. A coffee-break drill or a full loop.

🏢

Company track

Rehearse the actual loop: rounds, tone and focus areas per company. Five built in.

🎓

Campus

Placement-season mode: aptitude, coding and HR rounds, Indian campus style.

🎲

Blind mode

The AI picks the interview type, personality, difficulty and duration. You find out live.

Global

Sample questions

"Tell me about a time you took ownership of a difficult task."

"What tradeoff did you make in a technical project?"

"How would you design a notification system?"

Track data from docsFrom the docsVerbatim from the project's data/company-tracks.json.

localhost:1234 · setup

Interview setup: type, interviewer personality, focus areas

Manual setup: interview type, interviewer personality, focus areas, coding-intensive mode. real screenshotReal screenshotCaptured from the actual interface.

localhost:1234 · blind mode

Blind mode: the AI randomizes the interview's hidden parameters. real screenshotReal screenshotCaptured from the actual interface.

"You won't know if the interviewer is helpful or strict until they start speaking. Adapt on the fly!"

Pick a personality — or don't: Neutral Professional · Very Strict · Calm & Polite · Highly Helpful · Friendly Conversational. from docsFrom the docsPersonalities visible in the product's setup screen.

Privacy — the differentiator

Your rehearsal stays yours.

You're uploading your resume and stumbling through answers out loud. That deserves architecture that treats your data like it matters — because the design does.

🔑
The deployment key never leaves the server
GEMINI_API_KEY lives in the Express server's environment. It is never embedded in the client bundle — the release gate even scans the built bundle to prove it.
🎟
Voice runs on one-use ephemeral tokens
Live sessions connect with single-use tokens minted by the server. A token can't be replayed, and there's nothing long-lived to steal.
🗝
Bring-your-own-key is tab-scoped
Prefer your own key? It lives in sessionStorage for the current tab only — never localStorage, never a database, gone when the tab closes.
💾
History stays in your browser
Interviews, reports, skill profiles, in-progress sessions — all in your browser's storage. Active sessions checkpoint every five seconds so a reload recovers safely.
📤
Export everything, or erase everything
The dashboard exports all your reports as JSON — or wipes them. Canvases export and import independently. Local-first means you're never asking permission for your own data.
🚫
No account, no cloud sync, no telemetry
There is no authentication backend and no cloud sync. (The README is explicit: add authenticated, owner-scoped storage before enabling multi-device persistence.)
🛡
A hardened server in front of it all
CSP and security headers, bounded JSON bodies, request timeouts, and per-IP rate limits — 40 text requests and 12 live tokens per IP every 15 minutes.

Privacy architecture from docsFrom the docsStated in the project's README privacy, data-lifecycle and environment sections.

Gallery

The real thing, pixel for pixel.

Every capture below is the actual product — no mockups, no renders. Click any shot to view it full-size.

/interview · live

Live voice sessionreal screenshot

/dashboard · performance hub

Performance Hubreal screenshot · mock data

/dashboard · pressure & skills

Pressure & skill analyticsreal screenshot

/dashboard · resume intelligence

Resume Intelligencereal screenshot

/dashboard · strengths

Strengths & improvementsreal screenshot

setup · manual

Manual setupreal screenshot

setup · personality

Personality & typereal screenshot

setup · blind mode

Blind modereal screenshot

Setup

Running in four commands.

Requirements

🟢
Node.js 20+ · npm 10+
The stack is React + TypeScript + Express.
🔑
A Gemini API key with Live access
Voice interviews need the Live API. Everything else in the Practice Arena works without a key.
🌐
Chromium · HTTPS in production
Chromium-based browser for the best audio experience; browsers require a secure context for microphone access.

Get a Gemini key

No Google Cloud CLI required:

AI Studio

Open Google AI Studio's API keys page and sign in.

Create key

Choose "Create API key", select a Cloud project if prompted.

Connect AI

In InterviewCoach, choose "Connect AI" and paste the key.

Verify

Select "Verify and use key" — the app makes a small provider request before accepting it.

# install
$ npm ci

# configure — copy .env.example to .env, then set:
GEMINI_API_KEY=…

# start development mode
$ npm run dev
→ open http://localhost:1234

Environment

Variable	Default	Purpose
GEMINI_API_KEY	empty	Server-managed Gemini key
GEMINI_TEXT_MODEL	gemini-3.8-flash	Text generation, feedback, coaching
GEMINI_LIVE_MODEL	gemini-3.1-flash-live-preview	Real-time native-audio interview model
PORT	1234	Express listening port
AI_RATE_LIMIT	40	Text AI requests per IP / 15 min
LIVE_RATE_LIMIT	12	Ephemeral live tokens per IP / 15 min
VITE_REPOSITORY_URL	empty	Optional public repo link in the UI

Docker

$ docker build -t interviewcoach-ai .
$ docker run --rm -p 8080:8080 --env-file .env interviewcoach-ai
# open http://localhost:8080 — put TLS at the load balancer in production

The release gate

npm run check runs the full gate: TypeScript typecheck → behavior tests → production build → client-bundle credential scan → server smoke tests.

Setup reference from docsFrom the docsCommands, variables and defaults from the project's README.

Architecture

Three layers. No black boxes.

The browser does the work it can do safely; the server guards the key and mints tokens; Gemini does the talking. Note what's not here: no auth backend, no database, no cloud sync.

Client

Browser

React SPA — landing, interview, practice, dashboard, open-source routes
Web Audio + Gemini Live connection over an ephemeral token
Local resume parsing in a PDF.js worker
Local interview and practice persistence
JavaScript Worker runner for the Practice Arena

Server

Express

Security headers and per-IP rate limits
Text-generation proxy — POST /api/ai/generate, /api/ai/validate
One-use Live token minting — POST /api/live/token
Config and health — GET /api/config, /api/health

Provider

Google Gemini API

Native-audio live model for voice interviews
Text model for feedback, coaching and problem generation

Architecture from docsFrom the docsThe layer diagram from the project's README, restyled.

Troubleshooting

Stuck? Start here.

Set GEMINI_API_KEY on the server, or add a session key from AI settings. Confirm GET /api/config reports aiConfigured: true. Without either key, the local Practice Arena still works — but live interviews and generated feedback are unavailable, with transparent local fallbacks for practice coaching and design scoring.

Confirm the key has Live API access, that GEMINI_LIVE_MODEL is currently available in your region, that outbound HTTPS/WSS is allowed, and that the page is served over HTTPS. Live is a preview service — pin a tested model in production and watch Google's deprecation notices.

Allow microphone permission in site settings, select a working input device, speak during the two-second sample, and retry. The preflight's record-and-playback step exists precisely so you catch this before a session.

Scanned image PDFs contain no extractable text — use a text-based PDF or paste the resume text manually. Files are limited to 5 MB and 20 pages.

Not necessarily. A server key covers everyone on a shared deployment; otherwise each user can supply their own through AI settings, verified with a small provider request first. A user key stays in sessionStorage for the current tab and disappears when it closes.

Nowhere you didn't put it. Interview history, reports, skill profiles and in-progress sessions live in your browser. Export them as JSON or erase them from the dashboard. There is no cloud sync and no account — multi-device persistence would need authenticated, owner-scoped storage added first.

Troubleshooting from docsFrom the docsVerbatim guidance from the project's README troubleshooting section.

Open source

MIT. Fork it. Make it yours.

InterviewCoach AI is MIT-licensed and local-first by design. Run it for yourself, self-host it for your team, or contribute back.

🛣

Dedicated routes

Landing, Interview, Practice Arena, Progress dashboard, and Open Source — with SPA deep-link fallback for /interview, /practice, /dashboard and /open-source.

🤝

Contributing

See CONTRIBUTING.md in the repo. Security issues follow SECURITY.md — the project takes its own threat model seriously.

📦

Your data, portable

Reports export as JSON, canvases as JSON or SVG, everything erasable. Local-first isn't a slogan here; it's the storage layer.

Star on GitHub Run it locally →

Licensing & routes from docsFrom the docsLicense and route list from the project's README.

Walk into every interview already rehearsed.

Interviews are won before you walk in.

High-stakes, low-rep

Friends are awkward, coaches are expensive

Question banks don't know you

Talk. Get interrupted. Recover. Like the real thing.

Natural interruption

Live transcription

Pause, resume, reconnect

Hardware preflight

It read your resume. All of it.

Parsed locally, in your browser

Questions about your experience

Topic confidence, honestly shown