Pargot.ai

Open-source PM mock interviews you can run on your own laptop. Talk to a live AI voice interviewer while a real-time coach tracks your framework, ticks off what you've covered, and gives feedback after every answer — then get a scored summary at the end.

• 🎙️ Live voice interviewer — "Alex", powered by Gemini Live native audio over LiveKit. Asks, clarifies, probes follow-ups.
• 🧭 Framework tracking — 9 PM frameworks (CIRCLES, IMPROVE, STAR, AARRR, HEART, Issue Tree, Porter's…) with live stage progress.
• 💡 Real-time coaching cues + per-turn feedback in a side panel, without pausing the interview.
• 🎚️ Interview settings — seniority (PM → Group/Principal PM) and question types (design, improvement, estimation, root cause, behavioral, strategy).
• 📊 End-of-interview summary — score, strengths, improvements, next steps. 30-minute session cap.
• 🔐 Optional auth mode — Google sign-in, interview history, and daily quotas for hosted deployments. Off by default.

---

Quick start (laptop)

What you need

Tool	Version
Node.js	20+
pnpm	9+ (npm i -g pnpm)
Python	3.10+
uv	latest (install)

Plus two free API accounts:

• LiveKit Cloud project — cloud.livekit.io → Settings → Keys (free tier is plenty)
• Google AI Studio key — aistudio.google.com/apikey
• (optional) Groq key — console.groq.com/keys for a faster coach

No database, no OAuth, no account system needed to run locally.

1. Clone

git clone https://github.com/<you>/pargot.git
cd pargot

2. Voice agent

cd agent
cp .env.example .env          # fill in LIVEKIT_* and GOOGLE_API_KEY
uv venv
uv pip install -r pyproject.toml

3. Web app

cd web
cp .env.example .env.local    # fill in LIVEKIT_* and GOOGLE_API_KEY (same values as the agent)
pnpm install

4. Run (two terminals)

# Terminal 1 — agent
cd agent
source .venv/bin/activate     # Windows: .venv\Scripts\activate
python interviewer.py dev

# Terminal 2 — web
cd web
pnpm dev

Open http://localhost:3000, click Start interview, pick your settings, and allow microphone access.

Tip: if you see ModuleNotFoundError: No module named 'google.genai', you're running the system Python — activate agent/.venv first.

---

How it works

Browser (Next.js, web/)
  ├─ Voice room ───────────► LiveKit Cloud ──dispatch──► Python agent (agent/)
  │                                                       └─ Gemini Live native audio ("Alex")
  ├─ Coach panel ── /api/support, /api/feedback, /api/classify-turn
  │                  └─ Groq gpt-oss-120b → Gemini Flash → Flash-Lite → Gemma (auto-fallback)
  └─ Summary ────── /api/summary (Gemini 2.5 Flash)

• /api/token mints a LiveKit JWT and passes your interview settings to the agent via room metadata.
• The coach routes walk a model chain. If a model is rate-limited or overloaded it's marked down and the next one becomes the default; recovery is probed in the background, never while you wait.
• Frameworks live in a static catalog (web/lib/frameworks.ts); the model only emits IDs and flags, and the UI joins them.

Project structure

pargot/
├── agent/                    # Python LiveKit voice agent
│   ├── interviewer.py        # entrypoint
│   ├── prompts.py            # interviewer persona + question bank
│   └── settings.py           # reads seniority / question types from room metadata
├── web/                      # Next.js 15 app
│   ├── app/api/              # token, support, feedback, classify-turn, summary, interviews, auth
│   ├── components/app/       # interview room, coach panel, settings, summary, history
│   ├── hooks/                # coach + summary hooks
│   └── lib/                  # prompts, framework catalog, schemas, db, auth config
└── docs/coach-roadmap.md     # coach design notes

---

Auth mode (optional)

By default Pargot.ai runs with no sign-in, no database and no limits — ideal for personal use on your own machine. Don't expose that mode on a public URL: anyone could mint LiveKit tokens and spend your API quota.

For a public deployment, turn on auth mode in web/.env.local:

NEXT_PUBLIC_AUTH_ENABLED=true
AUTH_SECRET=...            # npx auth secret
AUTH_GOOGLE_ID=...         # Google OAuth client
AUTH_GOOGLE_SECRET=...
DATABASE_URL=...           # Postgres; Neon's free tier works well
DAILY_INTERVIEW_LIMIT=3
NEXT_PUBLIC_DAILY_INTERVIEW_LIMIT=3

Then create the tables:

cd web
pnpm db:push

This adds Google sign-in, a per-user interview history page (/history), and a daily interview quota enforced in /api/token. Google OAuth redirect URI: http://localhost:3000/api/auth/callback/google (plus your production domain).

---

Deployment

See DEPLOY.md — web on Vercel, agent on Railway (Docker), with env-var checklists.

---

Contributing

Issues and PRs are welcome. Before opening a PR:

cd web
pnpm lint
pnpm format:check
pnpm build

Never commit .env, .env.local, or any real keys — the .gitignore excludes them; the *.example files are the reference.

---

License

Pargot.ai is licensed under the GNU Affero General Public License v3.0. If you run a modified version as a network service, you must make your source code available to its users.

The web app began from LiveKit's agent-starter-react template (MIT); its original notice is kept in web/LICENSE-LIVEKIT-STARTER.