CourseMateAI
A study assistant that stays inside your course material, with five ways to learn from it.
- Role
- Agentic AI Engineer
- Timeline
- Jan – Feb 2026
- Team
- Solo
- Status
- Completed
The problem
Students drown in slides, papers and textbooks, and rote reading does not build understanding. General AI assistants hallucinate when asked about specific course material, which makes them unreliable for studying. A study tool has to stay grounded in what the student actually uploaded.
What it does
Grounded chat
Answers come from retrieved chunks of your own documents, with the model told to say so when the material does not cover a question.
Notes and flashcards
Turns documents into study notes and flashcards for active recall.
Quiz mode
Generates questions from your material and checks your answers.
Arena
Battle AI boss opponents that are generated through a two-stage LLM pipeline.
Structure-aware ingestion
Chunking that respects slide and section boundaries, with overlap and section-header metadata for better retrieval.
How it flows
- Upload slides and PDFs
- Structure-aware chunking
- FAISS index
- Mode-specific retrieval
- Grounded generation
- Auto-corrected JSON
Problems solved
Hallucination
ProblemLLMs produce plausible but wrong answers about specific course material.
SolutionStrict document grounding with retrieval, so answers stay tied to the uploaded content.
Five modes, five behaviours
ProblemChat, notes, flashcards, quiz and arena each need different retrieval and output formats.
SolutionLangGraph state machines with mode-specific nodes, prompts and output contracts.
Structured output that breaks
ProblemModels sometimes return malformed JSON, which would crash a quiz or flashcard screen.
SolutionA two-pass auto-correction step repairs the output before it reaches the UI.