Ujwol Shrestha

Case study

Final-year project · Full-stack AI application

AI Study Planner

AASA, the Adaptive AI Study Architect

Upload a lecture PDF and your exam date. Get topics, quizzes, answers you can check, and a study plan that adapts after every question.

Frontend

React app

Vite · 17 pages

  • Upload a lecture PDF
  • Quizzes and study sessions
  • AI chat with citations
  • Adaptive study planner
  • Analytics and reports

Backend

Spring Boot API

Java 17 · REST · JWT

  • Background PDF processing
  • Retrieval and hybrid re-ranking
  • Bayesian Knowledge Tracing
  • Weakness scoring
  • Plan re-prioritisation

Data

PostgreSQL 17 + pgvector

Relational data and 768-dim embeddings

AI

Gemini API

Embeddings and grounded answers

ML

Python · FastAPI

Random Forest weakness model

How a question gets answered

  1. Question
  2. Embed
  3. Top 20 by similarity
  4. Hybrid re-rank
  5. Best 5 passages
  6. Answer + citations
System architecture, drawn from the project repository.

The problem

Students get a stack of lecture PDFs and a fixed exam date. General AI chatbots can explain things, but they don't know your course material, they can make things up, and they have no idea which topics you're actually weak in. So they can't tell you what to study next.

What I built

A full-stack web app that turns a lecture PDF into an adaptive study programme. It extracts topics, generates multiple-choice quizzes, answers questions using only the uploaded material (with citations), tracks mastery per topic and keeps re-ranking a personal study plan until exam day. It also has accounts with email OTP and Google sign-in, analytics and reports.

How it works

  1. 01

    Uploads are processed in the background: text is extracted, split into ~512-token chunks and embedded with Gemini into PostgreSQL with pgvector.

  2. 02

    Questions use retrieval-augmented generation. Vector search pulls the top 20 passages, a hybrid re-ranker keeps the best 5, and Gemini answers with citations.

  3. 03

    Every quiz answer updates a Bayesian Knowledge Tracing estimate of how well the student knows each topic.

  4. 04

    A weakness score (mastery blended with a Random Forest model served from a Python service) plus the days left before the exam re-prioritise the plan.

Key decision

Answers grounded in the student's own material.

Instead of letting the model answer from general knowledge, every answer is built from retrieved passages of the uploaded PDF and cites them. Students can check an answer against their course notes instead of just trusting it.

Technology

  • React
  • Java 17 · Spring Boot
  • PostgreSQL + pgvector
  • Gemini API
  • Python · FastAPI
  • Docker

Have a problem to solve?

Something like this for your business? Let's talk.

Tell me what you're trying to achieve, or what you're still doing by hand. No technical brief needed.