I still remember the email. It was from a high school sophomore named Maria. She was using an early build of our AI tutor and was completely, hopelessly stuck on logarithmic functions. "It feels like I'm hitting a brick wall," she wrote. "The app keeps giving me the same type of problem, and I keep getting it wrong. I think I'm just not a math person."
That last line hit me hard. "Not a math person." What a terrible, limiting belief. And my product was reinforcing it. That night, I didn’t sleep. I just kept thinking about Maria. The problem wasn’t her; the problem was our code. Our "smart" tutor was acting like a dumb hammer, repeatedly hitting the same spot, expecting a different result. It was adapting, sure, but in the most basic, unhelpful way possible.
That email changed everything. It sent my team and me back to the drawing board and forced us to answer a much harder question: How do you build an AI that doesn’t just track right or wrong answers, but actually understands how a student learns?
The One-Size-Fits-None Problem
I’ve spent my career building companies and investing in founders who are trying to solve big problems. From scaling remote work with RemoteTeam to understanding media consumption at MovieLaLa, I’ve always been drawn to using technology to personalize experiences. But education felt different. It felt more important.
For over a century, our education system has operated on a factory model. A batch of students, all born in the same year, move down an assembly line at the same pace. A teacher, no matter how brilliant, has to teach to the middle, leaving some students bored and others, like Maria, completely lost. It’s a one-size-fits-none system.
We knew AI could be the answer. But the first wave of "adaptive learning" tools were, frankly, lazy. They were just glorified flashcard apps. They’d serve up a question. You get it right, you get a harder one. You get it wrong, you get an easier one. It’s a simple branching tree, not true adaptation. It doesn’t know if you’re struggling with the core concept or just made a silly mistake. It doesn’t know if you’re a visual learner or someone who needs to read the theory first. It has no real insight.
We wanted to build something better. Something that felt less like a machine and more like a world-class tutor sitting next to you. A tutor that could see the confusion on your face, adjust its explanation, and find the exact right way to make a concept click.
Data Is More Than Just Right or Wrong
The first step was to rethink our data. Most edtech platforms were only collecting one primary data point: the final answer. It’s a binary world of 1s and 0s. But the final answer is the least interesting part of the learning process.
We needed to go deeper. We started logging everything. And I mean everything.
- Response Time: How long does a student hesitate before answering? A long pause on a simple question is a huge red flag.
- Error Patterns: How did they get it wrong? Did they misapply a formula? Make a calculation error? Guess randomly? We built classifiers to categorize mistakes in real-time.
- Interaction Data: Did they re-read the question? Did they use the on-screen calculator? Did they look at the hint? Every click, every hover, every keystroke became a signal.
- Confidence Score: Before submitting an answer, we asked students to rate their confidence. This was a simple but powerful addition. A student who gets a question right but has low confidence hasn’t truly mastered the material.
Within three months, we had collected over 50 million of these granular data points. It was a firehose of information. We weren’t just seeing what students knew; we were seeing how they were thinking.
Building the "Brain"
With this rich dataset, we could finally build the core of our AI. We threw out the simple branching-tree model. Instead, we built a dynamic Bayesian network. That sounds complicated, but the idea is simple. For every student, the AI maintains a constantly updating map of their knowledge. Each concept (like "solving for x" or "understanding derivatives") is a node in the network. These nodes are all interconnected.
Think of it like a personal trainer for your brain. A good trainer doesn’t just make you do random exercises. They have a plan. They know that strengthening your core will help your squats. They know that improving your grip will help your pull-ups.
Our AI works the same way. It understands, for example, that a student struggling with quadratic equations might actually have a weakness in basic factoring. So, instead of hammering them with more quadratic problems, the AI will quietly slip in a few factoring exercises. If the student aces those, the AI updates its "knowledge map" and re-evaluates its strategy. Maybe the issue wasn’t factoring after all. Maybe it’s a problem with understanding the concept of a variable.
This is real adaptation. It’s not just adjusting the difficulty; it’s diagnosing the root cause of the struggle and dynamically changing the entire learning path. It’s the difference between a tool that tests you and a tutor that teaches you.
The Pizza Anecdote
I’ll never forget the day it all clicked. We were watching a session with a middle schooler who was having a terrible time with fractions. He just wasn’t getting it. The AI, based on his profile, noted that he was a highly visual learner who responded well to real-world analogies.
Our team had programmed a few of these analogies, but we were also experimenting with a generative component that could create its own. Suddenly, instead of showing another set of abstract circles and squares, the AI served up a simple animation of a pizza being sliced. It asked: "If you eat 2 out of 8 slices, what fraction of the pizza is left?"
The student, who had been failing for ten minutes straight, typed the correct answer in less than five seconds. The AI then followed up with another pizza question, then one about a chocolate bar, and then slowly removed the real-world objects to transition back to the abstract mathematical representation. He aced every single one. The AI had found his "on-ramp" to the concept.
That wasn’t a path we explicitly coded. It was a strategy the AI developed on its own by combining dozens of subtle data points: his past performance, his learning style profile, his hesitation time, and the success of similar strategies with other students. It was emergent intelligence. That’s when I knew we were building something more than just a product. We were building a partner in learning.
The Hardest Decision
Of course, it wasn’t all smooth sailing. Our biggest mistake was making the AI too good, too fast. In our quest for efficiency, our first major version was relentless. It adapted so quickly, constantly pushing the student to the very edge of their ability, that it was exhausting. Students felt like they were on a treadmill that was always speeding up.
We saw it in the data. Engagement was high for the first 15 minutes, but then it would fall off a cliff. Students were burning out. We had optimized for the algorithm, not the human.
It was a painful realization. We had to go back and intentionally make our AI less aggressive. We programmed in "plateau periods"—short breaks where the AI would serve up slightly easier problems to build confidence and let concepts sink in. We added more positive reinforcement and celebratory animations. It felt like a step backward. From a pure machine learning perspective, we were making the model less efficient. But from a human perspective, it was the most important change we made.
Building a truly great product isn’t about finding the perfect algorithm. It’s about finding the perfect balance between the machine’s logic and the user’s emotional experience. It’s a lesson I’ve learned over and over, from my first startup to my latest investment. You can have the best tech in the world, but if you ignore the human element, you’ve already lost.
Today, our AI tutor is used by thousands of students. We still get emails, but now they’re different. They’re from students like James, who finally understood calculus, or from parents who have seen their child’s confidence soar. And Maria? She’s now a freshman in college, majoring in computer science. She’s a "math person" after all.
Frequently Asked Questions
Do I need technical skills to built an ai tutor that adjusts as you learn?
Not necessarily. While technical understanding helps, the most important skills are clear thinking and the ability to break problems into smaller pieces. Many successful founders I've invested in started with zero technical background and either learned enough to be dangerous or found the right technical partner.
How long does it take to built an ai tutor that adjusts as you learn?
The timeline varies depending on your starting point and resources. For most founders, expect 2-4 weeks for initial setup and 2-3 months to see meaningful results. I've seen teams move faster when they focus on one thing at a time rather than trying to do everything at once.
What tools do I need to get started?
Start with the basics. You don't need expensive software or fancy tools. A spreadsheet, a note-taking app, and direct access to your customers will get you further than any enterprise platform. Add tools only when you hit a specific bottleneck.
How do I measure success with this approach?
Pick one or two metrics that directly tie to your goal and track them weekly. Vanity metrics like page views or follower counts rarely matter. Focus on metrics that reflect real engagement or revenue impact.