Interview Coach

Real-time AI commentary on interview answers. Listens to your mic, detects the interviewer's questions, streams per-question commentary from GPT-6 Luna, and lets you replay the whole session afterwards with timestamped notes.

Current version uses the browser's built-in Web Speech API for transcription (free, no extra API key needed). For a production deployment with longer sessions and better multi-language accuracy, swap lib/audioSession.ts for a Deepgram streaming implementation.

📘 Setting this up for the first time? See DEPLOY.md for step-by-step instructions on running locally, pushing to GitHub, and deploying to Vercel.

Architecture

Browser mic
    │
    ├─► MediaRecorder ─────────► Blob (saved for playback)
    │
    └─► Web Speech API ────► transcript stream
                                   │
                                   ▼
                     ┌─────────────────────────────┐
                     │       Orchestrator           │
                     │                              │
                     │  every final utterance →     │
                     │  POST /api/detect-question   │──► gpt-6-luna (is this a Q?)
                     │                              │
                     │  every ~220 chars of answer →│
                     │  POST /api/commentary (SSE)  │──► gpt-6-luna (streamed comment)
                     └─────────────────────────────┘
                                   │
                                   ▼
                            Zustand store
                                   │
                                   ▼
                             React UI

Setup

1. Install deps

npm install

2. Get a GPT-6 Luna API key

https://platform.openai.com/api-keys → Create new API key

3. Configure env

cp .env.local.example .env.local
# Edit .env.local and paste your key

4. Run

npm run dev

Open http://localhost:3000 in Chrome or Edge (Safari's Web Speech API is unreliable). Click Start, paste a JD, and start talking. First time, the browser will ask for microphone permission.

How it works

Live session flow

1. Start → Shows the JD + Resume modal.
2. Confirm → Opens the browser mic, starts Web Speech API + MediaRecorder.
3. Every finalized transcript utterance is sent to /api/detect-question. GPT-6 Luna decides if it's a new question. If yes → a new Question block appears at the top. If no → the utterance is added to the current Question's running answer buffer.
4. Every ~220 chars of new answer text triggers /api/commentary. This streams a one-sentence observation back via SSE; the frontend appends deltas as they arrive.
5. Pause stops the recorder and halts new commentary. Resume picks up where you left off.
6. End & Save stops everything, turns the recorded audio into an object URL, and pushes the session into Past Sessions.

Past session flow

Click any past session in the sidebar. You get:

• Full Q&A list with timestamps
• Audio player with question markers on the timeline (blue tick marks)
• Clicking a timestamp or marker jumps the audio there
• The currently-playing Q&A is highlighted

Data persistence

None. Sessions live in Zustand + memory only. Refresh the page and it's gone. This was a deliberate choice to ship faster — add IndexedDB or a backend DB later.

Tunables

In lib/orchestrator.ts:

• COMMENT_TRIGGER_CHARS (default 220) — how much new answer text triggers a comment. Lower = more frequent, chattier.
• COMMENT_MIN_GAP_MS (default 8000) — minimum time between comments on the same question, to prevent over-commenting during a long answer.

Known limitations (Web Speech API mode)

• Chrome / Edge only. Safari's implementation is too flaky to use.
• Engine disconnects every ~60s. The code auto-reconnects, but you may lose 1-2 words at the seam.
• Chinese accuracy is mediocre — especially with accent or background noise. If you want reliable Chinese, upgrade to Deepgram.
• No word-level timestamps — question markers on the playback timeline are based on when our app received the final utterance, not the true in-audio time. Usually within ~1s but drifts over long sessions.

Upgrading to Deepgram later

Replace lib/audioSession.ts with a streaming version (it needs to: open a wss://api.deepgram.com/v1/listen socket, pipe MediaRecorder chunks into it, and call the same four callbacks). Add a /api/deepgram-token route that mints short-lived project keys so the browser never sees your main Deepgram key. Everything else stays the same.