AI Question Paper Generator

Problem

Indian school exams have rigid structural constraints — specific board (CBSE, ICSE, state), class, subject, chapter coverage, controlled question-type and marks distribution — but teachers still draft them by hand. An LLM is well-suited to filling in question generation; the hard part is constraining it tightly enough to produce a board-shaped paper rather than a generic quiz.

Architecture

A hierarchical question bank (Boards → Classes → Subjects → Chapters) lives in MySQL and feeds prompt construction. The user picks a board, class, subject and chapter set, then specifies question type, difficulty and total marks. Gemini receives the constraints plus the relevant question-bank context; the response is parsed and rendered.

Output formats: PDF via WeasyPrint, Word via python-docx, with bilingual English / Hindi support (Noto Sans Devanagari embedded for Hindi). Answer keys are auto-generated with explanations.

Key engineering decisions

• Fallback chain: Gemini Flash → Pro → 2.0 Flash → local DB → placeholder. Any single model can rate-limit or fail. The chain degrades gracefully — Flash is fastest, Pro is the quality backstop, 2.0 Flash is the newer model, the local DB returns canned questions for the exact chapter, and the placeholder is the last-resort "paper still generates" guarantee.
• WeasyPrint for PDF, python-docx for Word. Schools need both formats. WeasyPrint handles multi-language typography correctly; python-docx is the only sane option for native .docx.
• Containerized via Docker, served with Gunicorn on Render. Keeps the deploy reproducible and the local dev environment identical to production.

Stack

Python 3, Flask, Gemini API, MySQL, WeasyPrint, python-docx, Noto Sans Devanagari, Docker, Gunicorn.

Links

• Live demo: api.simption.com/Ai_Intelligent_Paper_Generator/
• Repository: github.com/Anshuu2004/Ai_Intelligent_Paper_Generator