Gyaanify

The Idea

Any exam paper follows a structure — sections, questions, mark weights — but studying for one doesn’t. Students end up re-reading lecture slides or writing notes from scratch for each question. Gyaanify automates the note generation: feed in the exam paper, get back study notes calibrated to each question’s mark weight.

What It Does

Scanned PDFs go in. Per-question study notes come out.

The pipeline: OCR extracts raw text page by page, GPT-4-Turbo parses it into structured questions with section and mark metadata, then a five-stage LangGraph workflow generates notes for each question — context builder, first-pass generator (Gemini 2.5 Pro), teacher evaluator, clarity pass, and a final review. Results are stored in SQLite and exported to CSV.

The depth of each note is tied to the mark value: 1-mark questions get a tight 50-word definition, 15-mark questions get 700 words covering theory, subtopics, examples, and a recap.

The Build

Built in Python with LangChain and LangGraph. The multi-model setup was intentional — different stages have different requirements, so different models handle them. Gemini consistently produced the best first-pass notes on the subjects tested. GPT-4-Turbo handled structured extraction reliably at temperature 0.

The pipeline lives in Jupyter notebooks, with a standalone .py version for running end-to-end. A FastAPI stub exists but was never wired up.

Where It Stopped

The core pipeline works. What stalled it: the FastAPI layer (and the web interface beyond it) was more product work than the prototype warranted at the time. The more interesting engineering problems — fan-out parallel processing, quality scoring, caching repeated questions across paper years — were clear but not pursued.

Shelved in June 2025. The bench notes from the build are a better artifact than the code at this point.