AI EdTech
AI Lesson Agent
Upload any educational PDF. Get a personalized quiz where the AI structurally cannot give you the answers, even if you ask.
The Problem
Most AI tutors can be argued around.
No enforcement
"Tell me the answer" works. The same model that grades your response can be talked into skipping the question.
No isolation
The model holding the answer key is the same one giving hints. One well-phrased question is enough to leak it.
No ordering
Questions arrive in list order, not dependency order. Students get tested on advanced concepts before foundational ones.
This project tackles all three by making the problems structurally impossible, not by writing stricter prompts.
What Is Different
Construction guarantees, not prompt instructions.
Quiz cannot be skipped
No graph edge exists for skipping. The agent is a state machine. There is no code path from "question presented" to "next question" without a submitted answer.
Answers cannot leak to the hint path
Three role-isolated stages. The hint stage is constructed without the answer key in its context. Not instructed to avoid it. Never given it.
Questions follow prerequisite order
Concept relationships are stored as a graph in Neo4j. The quiz picks the question with the fewest unresolved dependencies first, a topological sort of what needs to be learned.
Questions are evaluated before you see them
Every question passes deterministic structural checks first — no meta-options, no duplicate choices, question must end with a question mark. Then a four-criteria binary evaluation scores it on answer unambiguity, distractor plausibility, objective alignment, and source grounding. All four must pass. Failures regenerate with a targeted critique, up to 3 attempts.
System Overview
Three role-isolated stages. One pipeline.
Built on LangGraph. Each agent has a fixed role and a fixed view of state. Context is scoped by construction, not by instruction.
How It Works
From upload to recap.
1
Upload your PDF
Drop any educational document. The system validates it is real content, not a blank scan, and extracts the text.
2
Review your learning plan
The Planner stage reads the document and proposes a set of learning objectives. You can edit, remove, or add objectives before approving. New objectives are validated against the document before approval.
3
The quiz begins
Questions are ordered by concept dependencies, foundations first. Each question passes structural checks and a four-criteria quality evaluation before you see it. Low-quality questions are regenerated with targeted feedback automatically.
4
Ask for a hint
Wrong answers unlock a hint from the Tutor stage, which was never given the answer key. It can point you in the right direction. It cannot tell you the answer.
5
Get your recap
When you finish, you get a summary of what you answered correctly on the first attempt, what you struggled with, and what concepts to revisit, ordered by prerequisite relationships.