ExamGenie

Turns your past exam papers into a personalized study plan — AI reads them, spots the topics that repeat, predicts what's next, then builds a mock exam and grades your answers.

Comment

Inspiration

Every student knows the ritual: before an exam, you dig up 5-10 years of past question papers and try to "spot" what's likely to come up. It works, but it's slow, manual, and mostly guesswork — you're pattern-matching by eye across dozens of PDFs and photos. We wanted to hand that pattern-matching to AI and turn hours of paper-flipping into a few clicks.

What it does

ExamGenie turns a stack of past exam papers into a personalized study plan. Upload photos or PDFs of previous question papers and it will:

  1. Read every question using AI vision (OCR) — no manual typing needed.
  2. Analyse the recurring topics across papers and rank them by frequency.
  3. Predict the topics most likely to show up in the next exam.
  4. Generate a realistic mock exam in the style of the real paper.
  5. Grade your answers with marks, feedback, and a model answer.

How we built it

  • Frontend: React + Vite for a fast, responsive upload-and-review flow.
  • Backend: Python + FastAPI serving a single API that wraps Google Gemini.
  • AI: Google Gemini (gemini-2.5-flash) via the official google-genai SDK handles everything — OCR on uploaded images/PDFs, topic-pattern analysis, mock-exam generation, and grading — all using structured outputs (response_schema) so the backend gets clean, typed JSON instead of having to parse free text.
  • Uploaded PDFs are split into small page chunks before being sent to Gemini, which keeps large multi-page past papers within request limits and makes progress visible instead of one long black-box call.

Challenges we ran into

  • Rate limits and large uploads. Several past papers at once, each several pages, can quickly hit Gemini's request limits. We built a chunking system (configurable pages-per-chunk and files-per-chunk) with retry/backoff logic so large batches degrade gracefully instead of failing outright.
  • Getting reliable structured data from a vision model. Free-text OCR output is hard to build a UI on top of. Using response_schema to force Gemini into consistent JSON shapes was key to making the frontend simple.
  • Keeping OCR accurate across messy scans. Handwritten notes, skewed phone photos, and low-quality scans all needed to resolve to the same clean question format.

Accomplishments that we're proud of

  • A single AI model (Gemini) powers the entire pipeline — OCR, analysis, generation, and grading — with no separate OCR service or ML model needed.
  • The whole flow, from raw photos of past papers to a graded mock exam, works end-to-end in under a minute.
  • Structured outputs mean the frontend never has to guess at parsing AI text.

What we learned

  • How far a single well-prompted multimodal model can go when you lean on structured outputs instead of stitching together multiple specialized services (OCR engine + NLP + grading logic).
  • Practical patterns for making LLM APIs resilient under real-world load: chunking, retries, and backoff matter as much as the prompts themselves.

What's next for ExamGenie

  • Support for more input types (typed notes, lecture slides) alongside past papers.
  • Multi-user accounts so students can track progress and predicted-topic accuracy across multiple exam cycles.
  • Deployment so judges/users can try it live without running it locally.

Built With

Share this project:

Updates