InterviewMinds

The Problem

Traditional mock interviews are expensive, hard to schedule, and existing AI tools lack depth. Enterprise organizations need secure, scalable candidate evaluation with real-time proctoring, collaborative coding, and comprehensive analytics — but building this requires integrating ML models, real-time systems, and production-grade infrastructure.

The Solution

Built a Turborepo monorepo with 81 backend service modules implementing 18 feature phases. Integrated Groq Llama 3.3-70b for sub-500ms AI responses, face-api.js with TensorFlow.js for real-time ML proctoring (face detection, eye tracking, expression analysis), a custom CRDT three-way merge algorithm for collaborative editing, WebRTC with STUN/TURN + optional SFU for video calls, and Piston API for sandboxed code execution in 20+ languages.

Key Challenges

• 1Implementing a CRDT three-way merge algorithm that preserves concurrent edits from multiple users without data loss — handling overlapping, non-overlapping, and append-type changes.
• 2Building an ML proctoring pipeline with graceful degradation — face-api.js for real inference when models load, content-derived analysis (never static values) as fallback.
• 3Preventing Redis quota exhaustion from BullMQ worker retry loops — implemented pause-on-error with exponential backoff.
• 4Designing a circuit breaker pattern for 5+ external services (Groq, Piston, Azure, Cloudinary) to prevent cascading failures.

Tech Stack