CareerCraft-AI

AI-powered career accelerator with a custom asymmetric RAG search engine, mock interview simulator, and gamified retention system.

React 19ViteTailwind CSS v4Framer MotionNode.jsExpress 5MongoDBMongooseOpenAI APIGroq APIJina AI EmbeddingsCloudinaryRazorpay
Live Demo →Source Code →

Problem

Job seekers juggle scattered tools for resume review, interview practice, and document Q&A — with no single platform that actually understands their documents.

Solution

Built a MERN platform with a custom-built asymmetric Retrieval-Augmented Generation pipeline (Jina embeddings + MongoDB Atlas vector search + Groq Llama 3.3) for grounded document chat, alongside AI mock interviews, flashcard/quiz generators, and a full gamification and subscription-tiering system.

Impact

Ships a production RAG pipeline with a lazy-evaluation parse cache (zero redundant Cloudinary calls after first open), strict JSON-constrained LLM outputs for quizzes, and a quota-enforced subscription system (Free/Pro/Enterprise).

Overview

CareerCraft-AI is an enterprise-grade AI career accelerator and document intelligence platform, built on a decoupled MERN architecture with a custom asymmetric RAG search engine, dynamic mock interview simulations, and gamified retention pipelines.

Core Features

  • Asymmetric RAG Pipeline: A recursive character splitter chunks documents at a 500-character window with 100-character overlap. Jina AI v3 uses two distinct embedding modes — retrieval.passage for indexed chunks and retrieval.query for user questions — so short queries align correctly with long document answers. MongoDB Atlas HNSW vector search runs cosine-similarity nearest-neighbor queries, pre-filtered by documentId so users never retrieve chunks from other documents.
  • Lazy Document Cache: Multer streams PDFs directly to Cloudinary on upload with zero local disk writes. pdf-parse runs once on first open and the raw text is cached to MongoDB — every subsequent summary, quiz, or chat request hits the cached field directly, with zero further network calls to Cloudinary.
  • Structured JSON Generation: Flashcards and MCQ quizzes use OpenAI's response_format: { type: "json_object" } constrained decoding, masking out any token that would break JSON syntax — guaranteeing clean parses with no markdown fencing.
  • AI Mock Interviewer: Interview settings (role, difficulty, company, skills) are injected into a detailed system prompt; full conversation history is reloaded from MongoDB and replayed to OpenAI each turn so the stateless API maintains contextual memory across the interview.
  • Gamification & Streaks: XP for logins, uploads, quizzes, and interviews, with a consecutive-day streak tracker, milestone bonuses, and daily XP caps to prevent abuse.
  • Subscription Tiering: Free/Pro/Enterprise tiers enforced via middleware that checks monthly usage counters (interviewsThisMonth, quizzesThisMonth) before every AI route, returning HTTP 403 once a limit is exceeded.

Tech Highlights

  • Frontend: React 19 on Vite with the Rust-based Rolldown bundler for near-instant HMR, Tailwind CSS v4's CSS-first engine, and react-markdown + remark-gfm to render AI responses as styled tables and syntax-highlighted code.
  • Backend: Express 5's native async error handling (no manual try/catch wrappers needed), JWT dual-cookie auth with OTP email verification via Nodemailer, and Razorpay-backed subscription billing with signature-verified webhooks.