I tried to build an AI that "understands" pedagogy. It was harder than I thought.

I’m Amir. I’ve spent the last few months building Edulavo, an AI platform that generates personalized learning roadmaps.

I wanted to share a technical struggle I faced, hoping for some feedback from other AI builders here.

The Challenge: "Hallucinated" Curriculums

When I first started with the MERN stack + LLMs, the AI would generate courses that looked good but made no sense. It would suggest learning "Advanced React Patterns" before "JavaScript Basics." It was just guessing keywords.

How I (mostly) fixed it:

Instead of one giant prompt, I broke the architecture into a pipeline:

1. The Assessor Agent: A dedicated prompt that only scores the user's chat answers (0-100 skill scale).

2. The Architect Agent: Takes that score and builds a JSON tree of modules (no content yet, just structure).

3. The Curator Agent: Only then searches our database/YouTube API to fill those modules with real content.

Separating the "Structure" from the "Content" reduced hallucinations by about 80%.

But... I need your eyes on it.

I’m still tweaking the "Curator Agent." Sometimes it pulls a video that is 5 years old. I’ve opened up a Free Plan where you can generate 2 full courses.

Could you try to generate a roadmap for a skill you already know? I want to know: Does the AI actually get the progression right, or is it still guessing?

Be as harsh as you want. I’m here to fix it.

https://www.edulavo.com/