We hit 50,000 users and all I could think was: we almost blew it.
Seriously. We were so close to building something nobody wanted. Back in the early days of our personalized learning platform, we were obsessed with the tech. We had this grand vision of an AI tutor, something we were calling 'Cognito', that could adapt to any student, any subject, anywhere. We spent a good six months holed up in a tiny office, whiteboarding algorithms and debating the merits of different NoSQL databases. You know, the classic startup engineering bubble.
Then we launched. And the first version was… a flop. A glorious, expensive, time-consuming flop. Engagement was terrible. Churn was through the roof. We had built a Ferrari engine, but we’d put it in a car with no wheels.
That failure was the best thing that ever happened to us. It forced us to stop acting like engineers and start acting like problem-solvers. It forced us to talk to actual students and teachers. It forced us to learn. Scaling to 50,000 users wasn't about the code; it was about the lessons we learned in the trenches. This is the stuff they don't teach you in business school. Here are the seven most important ones.
1. Personalization is a Lie (Almost)
Everyone in EdTech talks about "personalization." It's the holy grail. But here’s the dirty secret: most of it is just fancy content filtering. Showing a student a video instead of a text block isn't personalization. It's a preference setting.
We fell into this trap. Our first-generation AI was all about learning styles. Was this student a "visual" learner? A "kinesthetic" learner? We spent a year building a complex system to categorize students and deliver content that matched their supposed style. The result? It made zero difference. Absolutely none.
Real personalization isn’t about how you show the content. It’s about what you show, when you show it, and why. It’s about understanding a student's knowledge gaps at a granular level. So, we threw the whole learning styles engine in the trash. In its place, we built a 'knowledge graph.' Think of it like a giant spiderweb for, say, Algebra 1. We mapped every single concept—variables, equations, polynomials—and every dependency between them. Now, when a student struggles with, say, covalent bonds in chemistry, our system doesn't just show them another video. It checks if they understood the prerequisite concepts, like atomic structure and electron shells. It’s not about the format; it’s about the foundation.
We learned that you can’t personalize a learning path until you’ve depersonalized the knowledge itself. Break it down into the smallest possible atoms of information, then build the path back up for each individual. It’s way harder, but it’s the only thing that actually works.
2. The "Illusion of Choice" is Your Most Powerful Tool
We used to give students a massive library of content. "Look at all these resources!" we 'd proudly say. We thought more choice was always better. It’s not. It’s paralyzing.
Students would click around for a few minutes, get overwhelmed, and leave. They had decision fatigue before they even started learning.
The breakthrough came when we did the opposite. We drastically reduced the number of choices. Instead of a giant library, we gave them a single, recommended next step. Just one. "Here's the next concept you should work on."
But here’s the trick: we also gave them an escape hatch. A small, subtle button that said "Choose a different topic." Most students never clicked it. But knowing it was there gave them a sense of control. It created the illusion of choice. They felt like they were in the driver's seat, even though we were providing the GPS.
This is a powerful psychological principle. People want agency, but they also want guidance. By giving them a clear path and a subtle off-ramp, you satisfy both needs. Our engagement metrics tripled overnight when we implemented this. Don't drown your users in options. Give them a path and the feeling of control.
3. Engagement is About Agency, Not Entertainment
Another early mistake was trying to make learning "fun." We gamified everything. Points, badges, leaderboards—the whole nine yards. It was a disaster. It turned learning into a chore. Students were grinding for points, not knowledge. They were optimizing for the reward, not the understanding.
We were trying to compete with TikTok and Fortnite. That’s a game you can’t win. EdTech isn’t about entertainment. It’s about empowerment.
We stripped out all the superficial gamification. Instead, we focused on giving students a sense of agency and progress. We built tools that let them see their own knowledge graph grow in real-time. We engineered 'aha!' moments. We'd show them how solving for 'x' wasn't just a random skill, but the key that unlocked five other concepts they'd struggled with earlier. We made the progress itself the addiction. We made the progress itself the reward.
The most engaging feature we ever built was a simple dashboard that showed students their "knowledge frontier"—the boundary between what they knew and what they were ready to learn next. It was like a map of their own brain. They could see where they’d been and where they were going. That’s more powerful than any badge.
Stop trying to be fun. Start trying to be empowering. The motivation that comes from genuine understanding is far more sustainable than any extrinsic reward.
4. Data is a Compass, Not a Map
As a data-obsessed founder, this was a hard lesson for me. We had dashboards for everything. We tracked every click, every answer, every second of video watched. We thought the data would give us all the answers. We were looking for a map—a precise set of instructions on how to build the perfect product.
But data isn't a map. It’s a compass. It can tell you the general direction you’re heading, but it can’t tell you what the terrain looks like. It can tell you what is happening, but it can’t tell you why.
For months, our data showed a huge drop-off on a specific math lesson. We tried everything to fix it. We re-shot the video, rewrote the text, added more practice problems. Nothing worked. The data told us there was a problem, but it couldn’t tell us the cause.
Finally, we did something radical: we called the students. We talked to five of them who had dropped off at that exact point. Within 30 minutes, we had our answer. The lesson used a term that was defined in a previous unit, but the students had forgotten it. They were hitting a wall because they were missing a single piece of prerequisite knowledge. The fix wasn’t a better video; it was a simple link back to the definition. The data pointed us to the problem, but only human conversation could reveal the solution.
Don't get me wrong, data is essential. But it’s the start of the conversation, not the end. Use it to identify where to look, but then you have to actually go look. Get on the phone. Talk to your users. The most valuable insights are not in your dashboards.
5. The Ethical Tightrope is Real
When you’re building an AI that shapes how kids learn, you’re walking an ethical tightrope. And there’s no safety net. We had to confront some really tough questions.
What happens if the AI is wrong? What if it incorrectly assesses a student's knowledge and holds them back? Or pushes them forward before they’re ready? We’re not just talking about a bad recommendation on Netflix. We’re talking about a student’s educational future.
We also had to deal with bias. Our initial training data was sourced from a specific set of schools. What if our AI was implicitly optimizing for a certain type of student, leaving others behind? An AI is only as good as the data it’s trained on. And a lot of educational data has bias baked right in.
There are no easy answers here. For us, the solution was a "human in the loop" approach. We built tools for teachers to override the AI’s recommendations. We made the AI’s reasoning transparent, so a teacher could see why the system was suggesting a particular path. We treated the AI as a co-pilot, not an autopilot. The teacher is always the pilot in command.
This isn't a one-and-done fix. We started investing heavily in data diversity, actively seeking out data from underrepresented schools. We hired outside auditors to stress-test our algorithms for bias. It's a constant, expensive, and necessary process. You’re never "done" with ethics. It’s something you have to be actively working on every single day. If you’re in EdTech and you’re not having uncomfortable conversations about ethics, you’re doing it wrong.
6. Teachers Are Your Design Partners, Not Your Customers
For the first year, we barely talked to teachers. We were a direct-to-student platform. We thought we could bypass the school system entirely. That was arrogant and stupid.
We quickly learned that even if students are your users, teachers are your most critical partners. They are the ones on the front lines. They are the ones who see the students’ struggles and triumphs up close. They are the ultimate domain experts.
So we started a Teacher Advisory Board. We paid a dozen teachers—real, overworked, skeptical teachers—to come in and just destroy our product. And they did. It was brutal. One of them, a high school chemistry teacher named Mrs. Davison, told me flat out, 'This is a solution in search of a problem.' Ouch. They tore it apart. They pointed out all the flawed assumptions we had made. It was humbling, and it was invaluable.
They taught us that a learning platform isn’t just about the student and the content. It’s about the classroom context. How does this tool fit into a teacher’s workflow? How can it save them time? How can it give them the insights they need to intervene with a struggling student?
We shifted our focus from replacing teachers to empowering them. We built dashboards that gave them a real-time view of their entire class. We created tools that helped them identify which students needed help, and with what. We stopped trying to be the teacher and started trying to be the teacher’s best assistant.
Your product will be 10x better if you co-design it with the experts. Don’t treat them like a focus group. Treat them like a core part of your product team.
7. Build for the B-Minus Student
It’s tempting to build for the extremes. You either focus on the struggling students who need remediation, or the advanced students who need acceleration. We did both. And we ended up serving neither of them well.
The vast majority of students are in the middle. They’re the B-minus or C-plus students. They’re not failing, but they’re not excelling either. They’re just… getting by. They are the most underserved population in education. And they are your biggest opportunity.
These are the students who have small, compounding knowledge gaps. They missed a key concept in 6th-grade algebra, and it’s been haunting them ever since. They don’t need a full-blown intervention. They just need someone to find and fill that specific gap.
When we shifted our focus to this group, everything clicked. Our knowledge graph approach was perfectly suited to find those tiny, hidden gaps. We could deliver a 10-minute intervention that made a huge difference. For these students, it was a revelation. They weren’t "bad at math." They just had a few holes in their foundation.
Building a company is exactly like that. You start with a perfect, elegant vision on a whiteboard, and then the real world walks in and punches it in the face. All that time I spent angel investing in companies like Anthropic and Scale AI, I saw it from the outside. But when it's your own company, you feel it. Success isn't about being the smartest person in the room with the most brilliant ideas. It's about how fast you can admit you were wrong. It's about your learning velocity. We didn't get to 50,000 users because we were geniuses. We got there because we were willing to be humbled, and we never stopped learning.
Frequently Asked Questions
Can I implement all of these at once?
I'd strongly recommend against it. Pick the 2-3 items that resonate most with your current situation and focus there. Trying to do everything simultaneously is a recipe for doing nothing well.
How do I know which items apply to my situation?
Start by honestly assessing where your biggest bottleneck is right now. The items that address that specific constraint will give you the highest return on your time and energy.
Which item on this list has the highest impact?
It depends on your stage and context, but in my experience, the items near the top of the list tend to have the broadest applicability. That said, sometimes the less obvious items create the biggest breakthroughs for specific situations.
Are these recommendations still relevant in 2026?
Absolutely. While specific tools and tactics change, the underlying principles remain consistent. I update my thinking regularly based on what I'm seeing in the market and across my portfolio companies.