An AI tutor that helps you get unstuck without giving the answer away.

Brilliant teaches by doing. Getting stuck breaks that.

This is a self-initiated concept made for my portfolio. It is not affiliated with or endorsed by Brilliant.org.

Brilliant replaces passive lectures with hands-on puzzles that build intuition for math, science and computer science. This concept builds on their brand refresh and explores a shift from static help to an agentic learning experience: an AI that acts like a patient tutor.

• –Learners lose confidence when they get stuck with no guidance.
• –Static hints don’t know where the learner’s logic broke.
• –Once momentum breaks, many learners leave the lesson.

Too little help and learners quit. Too much and they stop learning.

Learning science calls the useful kind of difficulty productive struggle. The goal was to protect it. Help had to be sized to the exact gap in the learner’s thinking.

Learners don’t want answers. They want a nudge in the right direction.

Learners often repeat the same mistake without knowing why. But handing over the answer kills the sense of achievement. The insight was simple: nudge, don’t tell.

Meet Sam, the motivated learner

Sam is curious and independent. They love the “aha” moment of figuring something out alone, but need a lifeline when a logic gap stops their progress.

1. 01

   Sam needs

   To build confidence through solving problems on their own.

   So Koji will

   Offer the smallest useful hint first, never the answer.

2. 02

   Sam needs

   Help that understands their specific mistake.

   So Koji will

   Read attempts and time spent to find the real logic gap.

3. 03

   Sam needs

   Control over when and how help appears.

   So Koji will

   Make every nudge optional and easy to dismiss.

An agent that watches, thinks, and only then steps in.

Here, agentic means the AI acts on its own signals. It watches progress quietly and offers help at the moment a learner is about to hit a wall, not before.

Help escalates one rung at a time

Koji never jumps to the answer. Each level of help is only offered if the one before didn’t work, which keeps the learner doing as much of the thinking as possible.

Five design principles

• –Preserve productive struggle.
• –Step in selectively, not constantly.
• –Be transparent about why a hint appears.
• –Keep learners in control of help.
• –Support learning, not just completion.

Koji: a tutor built into the lesson, not bolted on beside it.

Instead of a separate chatbot, Koji lives inside the lesson. It uses real-time signals to offer adaptive guidance and targeted practice. I designed three key moments, each defined by its trigger, the AI’s behavior, and where the learner stays in control.

Flow 1: Getting unstuck

• –Trigger: hesitation or repeated wrong attempts.
• –AI behavior: Koji offers an optional nudge toward the key principle.
• –Learner control: accept the nudge or keep trying alone.

Flow 2: Chat with Koji

• –Trigger: the learner asks Koji for deeper help.
• –AI behavior: Socratic questions that move from concept hints to specific examples.
• –Learner control: they decide how far up the help ladder to go.

Flow 3: After the lesson

• –Trigger: a lesson or module is complete.
• –AI behavior: Koji summarizes progress and suggests personal next steps.
• –Learner control: suggestions are options, not instructions.

How I would know if it works.

This is a concept, so it hasn’t been tested with learners yet. Here is how I would measure it. The North Star metric: puzzles completed after a learner gets stuck.

Is it helping learners?

• –Momentum: do learners get past hard puzzles faster?
• –Completion: do more learners finish the hardest levels?
• –Retention: does the help lead to better scores in later lessons?
• –Confidence: do learners feel more capable, not more dependent?

Is the AI behaving well?

• –Timing: does help arrive at the right moment?
• –Restraint: does it avoid giving away answers?
• –Control: do learners feel in charge of when they get help?

The hard part isn’t making AI smart. It’s knowing when to stay quiet.

This project showed me that good AI UX is often about restraint. The interesting design work was deciding when Koji should step in and when to let the learner think.

What I would do next:

• –Test with real learners to tune the timing of each nudge.
• –Make Koji explain why it offered a specific hint.
• –Explore how support should change over weeks, not just one lesson.