I built an AI that turns any question into a full 3-part lesson and it's powered by Gemma 4

Hey everyone,
I want to talk about a problem that I think most of us have quietly accepted as normal: we ask AI a question, we get an answer, and within minutes it's gone. Not because the answer was wrong. Because we never had to do anything with it. We just consumed it, the way you consume a snack you forget you ate.
That gap between "getting an answer" and "actually learning something" is what pushed me to build RealLearn.
The core insight
Think about how real learning has always worked, in a good classroom, from a good mentor, in any subject that ever actually stuck with you. It's never one paragraph. It's a sequence. You get the foundation first, so you have something to stand on. Then you get the mechanism, the part that explains why it works and not just that it works. Then, if the teaching is good, someone connects it to something real, something happening right now, so the idea stops being abstract and starts being useful.
Most AI tools skip straight past all of that and hand you a summary. RealLearn tries to rebuild the whole sequence, automatically, for any question you throw at it.
How it actually works
You ask something, anything from "how do black holes work" to "why did inflation spike in 2024," and RealLearn builds a three-part journey:

Foundation. A clear, beginner-friendly framing of the idea. No jargon dump, no assumption that you already know half the vocabulary.
Mechanism. The deeper layer. The "how and why" that actually explains the thing instead of describing it.
Real World Now. The idea connected to current events, pulled from live news, with real names, dates, and numbers. This is the part that makes a concept feel less like trivia and more like something worth knowing.

Here's the part I care about most: you can't just scroll past these sections. Each one is gated by a short quiz. Get it right, and the next part unlocks. Get it wrong, and you're not just told the correct answer, you get a full explanation folded into the feedback, so even your mistakes end up teaching you something. Answer choices get reshuffled on retakes too, so you can't just pattern match your way through.
Everything adapts. You choose your language (eight of them, generated natively rather than translated after the fact, so nuance and tone survive the trip), your level (from middle school through college and beyond), and your pace (a fast mode for quick hits, an "Explain" mode for the full journey).
What's under the hood
None of this works if the AI output is flaky, so a good chunk of the engineering effort went into reliability rather than flash. Lessons are generated by Gemma 4 (gemma-4-26b-a4b-it) through Cloudflare Workers AI, wrapped in a system that enforces strict structure, repairs malformed output when the model stumbles, validates everything before it ever reaches your screen, and gracefully degrades instead of breaking outright when something goes wrong. There's a caching layer to keep things fast, and a moderation layer to keep things safe.
None of it is visible to you as a user. It's just supposed to feel calm, fast, and dependable. Whether it does is really the question I'm here to ask.
Where this came from
RealLearn started as my submission to the Gemma 4 Good Hackathon on Kaggle. It's grown a fair bit since then, but the original motivation hasn't changed: I wanted something that treats a question as the start of a small education, not the end of a search.

What I'm actually curious about
Does gating progress behind a quiz genuinely help you retain something, or does it start to feel like busywork after the second or third question? I have my own theory, obviously, since I built the thing, but I'd rather hear it from people who didn't.