Professor Mind Hack

outlook of the project

Inspiration

RUET students always say things like "Rahman স্যার loves numerical questions" — but this knowledge is scattered word-of-mouth, lost every batch. We wanted to turn it into real data. Generic "chat with PDF" tools treat every paper in isolation; they can't tell you a topic repeated across 4 years. We wanted a tool that understands a professor's pattern, not just one paper.

What it does

Upload scanned question papers → Gemini extracts every question (topic, type, difficulty, marks) → a deterministic engine aggregates this across years to reveal topic weightage, numerical-vs-theory ratio, and recurring topics, each with a confidence score. It can even generate a practice exam paper matching a professor's historical style.

How we built it

React + TypeScript + Tailwind + Recharts frontend, Express/Node backend with the @google/genai SDK. Gemini 3.7 Flash reads scanned papers directly (no separate OCR) and returns strict, Zod-validated JSON. All statistics — frequency, recurrence, confidence — are computed with deterministic code, not AI, keeping the numbers trustworthy.

Challenges we ran into

Getting Gemini to return consistently structured JSON from messy scans took prompt/schema iteration. We also had to avoid overconfident predictions from thin data, which is why we built confidence scoring in from the start. Wiring the full pipeline — upload → extract → validate → aggregate → visualize — within a few hours meant ruthlessly prioritizing a working core over extra features.

Accomplishments that we're proud of

A real multimodal pipeline working end-to-end, not a mockup
Multi-year pattern analysis that generic doc-QA tools can't do
Confidence scoring so the tool never overstates thin data
A working exam-paper simulator based on real professor patterns

What we learned

Schema design matters as much as prompting when using an LLM for structured extraction. Splitting responsibilities — AI for reading unstructured documents, code for statistics — made the app both more trustworthy and easier to debug fast.

What's next for Professor Mind Hack

Crowdsourced uploads across departments, teacher-vs-teacher comparison, personalized study plans from weak topics, better OCR on low-quality scans, and expanding to lecture slides and lab manuals.