LiveKit AI Interview Agent — Node.js

A real-time AI-powered interview agent built on LiveKit Agents SDK (Node.js 1.x). The agent conducts structured technical interviews using voice, with automatic transcription, topic tracking, and MongoDB integration.

Architecture

┌─────────────────┐    ┌──────────────────┐    ┌─────────────────┐
│   Frontend      │───▶│   Token Server   │───▶│  LiveKit Server │
│ (HTML/JS/CSS)   │    │  (Express, 8081) │    │   (ws://7880)   │
└─────────────────┘    └──────────────────┘    └────────┬────────┘
                                                        │
                                                        ▼
                                               ┌─────────────────┐
                                               │   Agent Worker   │
                                               │  (agent-node/)   │
                                               │                  │
                                               │  STT: Sarvam AI  │
                                               │  TTS: Sarvam AI  │
                                               │  LLM: Gemini 2.5 │
                                               │  VAD: Silero     │
                                               └────────┬────────┘
                                                        │
                                                        ▼
                                               ┌─────────────────┐
                                               │    MongoDB       │
                                               │  (Transcripts)   │
                                               └─────────────────┘

Project Structure

├── agent-node/              # LiveKit Agent Worker (Node.js)
│   ├── agent.js             # Main entry point — session lifecycle
│   ├── prompts.js           # System prompt builder (resume + MongoDB job config)
│   ├── tools.js             # LLM tools (transition_topic, end_call)
│   ├── logger.js            # Transcript file writer with structured markers
│   ├── processor.js         # Post-interview MongoDB upload
│   └── package.json
│
├── token-server/            # Express server for LiveKit token generation
│   ├── server.js            # Token endpoint + resume upload + static serving
│   └── package.json
│
├── frontend/                # Browser-based interview UI
│   ├── index.html           # Interview page
│   ├── app.js               # LiveKit SDK client logic
│   └── style.css            # Styling
│
├── .env.example             # Environment variable template
├── .gitignore
└── README.md

Prerequisites

• Node.js ≥ 18
• LiveKit Server — Self-hosted or cloud (docs)
• API Keys:
  • Sarvam AI — STT (saaras:v3) & TTS (bulbul:v3)
  • Google AI — Gemini 2.5 Flash
• MongoDB — For job configs and transcript storage

Getting Started & Initialization

Follow these steps to initialize and run the project from scratch.

1. Prerequisite Installations

Ensure you have the following installed on your system:

• Node.js (v18 or higher)
• MongoDB (Local or Atlas connection string)
• Git

2. Clone the Repository

git clone <your-repository-url>
cd <your-repository-directory>

3. Install Dependencies

You must install Node modules for both the agent worker and the token server.

# Install agent dependencies
cd agent-node
npm install
cd ..

# Install token server dependencies
cd token-server
npm install
cd ..

4. Configure Environment Variables

Create your configuration file from the template:

cp .env.example .env

Open .env in a text editor and populate your actual keys:

• SARVAM_API_KEY (For STT & TTS)
• GOOGLE_API_KEY (For Gemini 2.5 Flash)
• MONGODB_URI (Your MongoDB connection string)
• (LiveKit keys can be left as default dev keys if using local server)

5. Download LiveKit Server Binary

If you are running LiveKit locally, download the pre-compiled binary for your OS from the LiveKit GitHub Releases page. Place the livekit-server.exe (or equivalent binary) directly into the root directory of this project.

6. Start the Project Services

A convenience batch file is provided for Windows users to launch all required services simultaneously.

Option A: Using the Batch script (Windows) Simply double-click the start.bat file in the root directory. This will automatically open three separate command prompt windows running:

1. LiveKit Server (in --dev mode)
2. Token Server
3. Agent Worker Node

Option B: Manual Startup (Mac/Linux/Windows) Open three separate terminal sessions in the root directory:

Terminal 1 (LiveKit):

./livekit-server --dev

Terminal 2 (Token Server):

cd token-server
npm run dev

Terminal 3 (Agent Worker):

cd agent-node
node agent.js dev

7. Open the Application

Once all three services are running successfully, open your browser and navigate to: http://localhost:8081

How It Works

Interview Flow

1. User uploads resume → Token server stores PDF and creates a LiveKit room with metadata
2. Agent joins room → Reads room metadata (resume path, job ID)
3. System prompt built → Parses resume PDF + fetches job config from MongoDB
4. Voice interview begins → Agent greets candidate and follows topic structure
5. Topic tracking → transition_topic tool logs [TOPIC_START/END] markers with skills & proficiency levels
6. End call → end_call tool logs [SESSION_END], says farewell, and triggers shutdown
7. Post-processing → Transcript parsed and uploaded to MongoDB with structured document

Transcript Format

[SESSION_START] 2026-03-31T16:03:18.000Z | Job: Senior React Developer
[METADATA] CAND:user123 | APP:app456 | JOB:69cbbe0730202aca28dd4281
[SKILLS] React:L4, JavaScript:L4, TypeScript:L3, HTML/CSS Development:L3
[TOPICS] Frontend Development & UI Implementation, Type Safety & Application Architecture

Interviewer: Hello! My name is Ritu, and I will be your interviewer today.
Candidate: Hi Ritu, my day has been good.

[TOPIC_START] Frontend Development & UI Implementation | React, JavaScript, HTML/CSS Development | L4 | 2026-03-31T...
Interviewer: Could you describe your experience developing responsive web applications?
Candidate: Sure, at my previous role I built...
[TOPIC_END] Frontend Development & UI Implementation | React, JavaScript, HTML/CSS Development | L4 | 2026-03-31T...

[SESSION_END] 2026-03-31T16:15:00.000Z

MongoDB Document Schema

{
  "candidate_id": "user123",
  "application_id": "app456",
  "job_id": "69cbbe07...",
  "conversation_log": "Interviewer: Hello!...\nCandidate: Hi...",
  "topics": [{ "name": "Frontend Dev", "skills_based_on": ["React", "JS"] }],
  "skills": ["React:L4", "JavaScript:L4"],
  "topic_logs": [{
    "topic": "Frontend Dev",
    "skills": ["React", "JS"],
    "start_time": "2026-03-31T...",
    "end_time": "2026-03-31T...",
    "log": "Interviewer: Could you...\nCandidate: Sure..."
  }],
  "created_at": "2026-03-31T...",
  "status": "completed"
}

AI Components

Component	Provider	Model	Purpose
STT	Sarvam AI	saaras:v3	Speech-to-text (Hindi/English)
TTS	Sarvam AI	bulbul:v3 (ritu)	Text-to-speech
LLM	Google	gemini-2.5-flash	Interview responses & tool calls
VAD	Silero	—	Voice activity detection

Environment Variables

Variable	Description
LIVEKIT_URL	LiveKit server WebSocket URL
LIVEKIT_API_KEY	LiveKit API key
LIVEKIT_API_SECRET	LiveKit API secret
SARVAM_API_KEY	Sarvam AI API key (STT + TTS)
GOOGLE_API_KEY	Google Gemini API key
INTERVIEW_API_KEY	Token server auth key
TOKEN_SERVER_PORT	Token server port (default: 8081)
MONGODB_URI	MongoDB connection string
INTERVIEW_DB	Database name
TRANSCRIPT_COLLECTION	Collection for transcripts
JOB_COLLECTION	Collection for job configs

Deployment Notes

• The agent worker runs as a persistent process and auto-accepts job requests from LiveKit
• Resume PDFs are stored in resumes/ (gitignored) and cleaned up by the token server
• Transcripts are written to transcripts/ during interviews and deleted after MongoDB upload
• For production, use LIVEKIT_URL=wss://your-livekit-domain.com with proper TLS