AI Adaptive Interview Platform

A high-performance, full-stack AI interview platform featuring real-time video monitoring, behavioral analysis, and automated reporting.

🚀 System Architecture

The platform is built with a modern, decoupled architecture designed for low latency and high scalability.

🏗️ Core Components

Component	Technology	Role
Frontend	HTML5, CSS3, JavaScript (ES6+)	Candidate interface & Admin dashboard.
Backend	FastAPI (Python 3.11)	REST API & WebSocket server.
Real-time Engine	Socket.IO	Low-latency binary video streaming & metadata sync.
Database	MongoDB Atlas	Persistent storage for interview data & analytics.
Cloud Storage	Cloudinary	Permanent storage for full interview recordings.
AI/ML	OpenAI / OpenRouter	Automated answer analysis & behavioral insights.

🔄 System Workflow

1. Session Initiation: Candidate joins via a unique link; the system initializes camera/mic and establishes a WebSocket connection.
2. Live Streaming:
   • Video: Captured via <video> + <canvas>, throttled via requestAnimationFrame, and sent as binary JPEG blobs via Socket.IO.
   • Metadata: Transcripts, audio levels, and proctoring alerts are synced in real-time.
3. Admin Monitoring: Admins receive binary frames and metadata instantly, rendering them efficiently using URL.createObjectURL to minimize DOM overhead.
4. Interview Completion: Full session recording is stopped, packaged as a .webm blob, and uploaded to Cloudinary.
5. Persistence: Cloudinary URLs and interview analytics are saved to MongoDB Atlas for long-term access.

🛠️ Issues Solved & Implementation Details

1. Live Preview Inconsistency & Latency

• Issue: Originally used HTTP polling (setInterval + fetch), resulting in 3-5s delays and high server overhead.
• Solution: Implemented Socket.IO with binary frame streaming.
• Why: WebSockets provide a persistent, full-duplex connection. Binary JPEG blobs (320px width, 0.4 quality) significantly reduce payload size compared to Base64 strings, ensuring <150ms latency.

2. Video Persistence

• Issue: Videos were saved to the local server filesystem, which is ephemeral on platforms like Render/Vercel, leading to data loss.
• Solution: Integrated Cloudinary for cloud-native video storage.
• Why: Cloudinary provides reliable, permanent storage with a global CDN, ensuring interview recordings are never lost and are easily accessible via secure URLs.

3. Performance Optimization

• Issue: High CPU usage on the candidate side due to frequent frame capture.
• Solution: Replaced setInterval with requestAnimationFrame and implemented a 120ms throttle (~8 FPS).
• Why: requestAnimationFrame aligns with the browser's refresh rate and pauses when the tab is inactive, saving resources while maintaining a smooth visual experience for the admin.

4. Distributed State Persistence (Critical Fix)

• Issue: The backend used an in-memory dictionary (interviews = {}) to track session state. This caused "Interview not found" errors when the server restarted or when multiple worker instances were used (common on Render/Vercel).
• Solution: Migrated all session state to MongoDB Atlas.
• Why: By using a centralized database for state, the application becomes stateless and horizontally scalable. Sessions persist across server restarts, and any worker instance can handle any request, ensuring 100% session reliability.

5. Deployment & Initialization Safety

• Issue: The backend would crash on startup if Cloudinary was not properly imported or if environment variables were missing.
• Solution: Fixed import ordering and added safe initialization guards.
• Why: Ensures the backend starts gracefully even if external services are temporarily misconfigured, providing clear warning logs instead of a total system failure.

⚙️ Setup & Deployment

Environment Variables

MONGO_URI=your_mongodb_uri
CLOUDINARY_CLOUD_NAME=your_cloud_name
CLOUDINARY_API_KEY=your_api_key
CLOUDINARY_API_SECRET=your_api_secret
OPENROUTER_API_KEY=your_api_key

Installation

pip install -r requirements.txt
python backend/uploded.py