2025-11-03
Tagged: personal, cartesian tutor, strategy
6 months ago, I started building Cartesian Tutor to explore AI tutors as the future of education. It’s been an incredibly rewarding journey and despite the lack of commercial success, I’ve learned a tremendous amount. A++ experience, highly recommended.
I’ll be job searching over the next few months. If you are looking for an AI-pilled staff software engineer with deep expertise in ML and LLMs who is also capable of fullstack prototyping and thinking holistically about product/technology/strategy, let me know. My search pipeline is currently underindexed on Boston-area / remote jobs, so these are especially welcome.
I’m happy to keep Cartesian Tutor up and running for students’ sake, and as a continuing hobby project.
In this essay, I’ll do a light postmortem and talk about what went well, what went poorly, and how I would do things differently next time.
Cartesian Tutor scales 1-on-1 tutoring with AI.
Education is due for a change. The standard classroom model has many weaknesses: too many students to have dedicated 1:1 time for each student, too few students to achieve economies of scale in producing quality content, and enough variation in ability that the entire class moves at the pace of its 25%ile student.
Textbooks and YouTube creators set the bar for scalability. 1:1 tutoring sets the bar for effectiveness. Could we have both with AI tutors?
Let’s look at why tutors are so effective. A tutor builds a mental model of their student: thinking style, misconceptions, knowledge gaps, working habits, and more. Using this mental model, the human tutor plans and delivers the specific lesson that most rapidly brings the student to mastery. Broken down as subtasks:
The Socratic method built into ChatGPT’s Study Mode is but one tiny step in this direction. I found that LLMs were not particularly good at any of the other subtasks. (That hasn’t stopped a number of AI education companies from pushing slop generation products to teachers…)
Cartesian Tutor uses a mix of AI-delivered lessons/problem review with traditional software and hand-curated curriculum, practice tests, and lesson plans, and the result seems to work decently. An especially important ingredient here is human exam-writers’ taste in writing good questions that can’t just be solved by pattern matching/plug-and-chug. Students’ failed attempts at solving these olympiad questions create a trail that AI can follow to diagnose students’ weaknesses.
In addition to this core product offering, I also built:
When I started, I and many other people believed that there was some nonzero chance that AI would quickly overtake many job descriptions, empowering small teams to compete toe-to-toe with larger incumbents by leveraging AI agents. I tried my hand at setting up an AI council to dispense startup advice; using AI tools to generate marketing copy; using AI tools to vibecode my frontend. None of this panned out. I believe now that most jobs are safe, and “startup founder” the safest of all, given how much adaptability, taste, and judgment it demands.
So, what should we make of the many recent examples of highly successful AI-centric startups, all of which have shown unprecedented growth rates and revenue numbers? VCs would love to spin the narrative that “AI changes everything”, but I attribute this wave of hypergrowth startups to a different combination of factors:
So fundamentally, I am heading back into the job market because my assessment is that I can create (and capture) the most value by working as an AI specialist employee at one of these highly successful AI-centric companies, rather than as a generalist startup founder.
That being said, I think it is highly likely that I will try another startup in the future. Hence, the notes.
My efforts over these past 6 months were split between:
While I failed at my first goal, I made amazing progress on the second and third goals. I don’t regret learning about AI or building companies, but next time will be a clear focus on building a successful business.
I also spent some time consulting for another ed tech company, which generated some revenue, helped sharpen my consulting skills, and helped me see a very different way of working with Claude – but ultimately I think this was just another distraction I could do without next time.
I originally thought it would be a waste of money to go for a coworking subscription when I had a nice WFH setup, but in retrospect it would have been a good idea. My most productive times were in busy coffeeshop environments, even if I was missing my split keyboard/widescreen monitor.
Physical health could have used more attention. My previous daily routine included a bike ride into the office, and without that, I found my physical fitness gradually dropping, until I suffered some sort of acute lower back injury. I’ve recovered and am doing more yoga to help strengthen my core.
Overall, I was surprised at how often much I enjoyed random meetings with old and new friends. Part of this was undoubtedly the social isolation that comes with being a solo founder. I also found the conversations helpful in refining and developing ideas for my startup. This time, I spent ~2% of my time meeting people, and next time I think I should spend more like 5-10% meeting people. As a Boston-based founder, the base rate of fortuitous encounters is a lot lower, so it’s worth being deliberate about finding and chatting with people.
The weekly blogging thing was honestly great. It helped me sort through the zillion thoughts that were running through my head and set weekly goals for myself. Community-wise, many people I chatted with had read many or all of my updates. It sparked many good conversations, and at a time when so many people are looking for informed takes on where AI is going, it was a great way to build some reputational currency.
I worked roughly 30-50 hours/week during this startup period. Looking at the pattern of my work hours, I found myself working in extremely productive 2-4 hour bursts, followed by a few hours after to recover. I found the following times particularly productive
Given how little overlap there is here with a traditional 9-5 working schedule. I’ll have to think about how I can create these optimal conditions in my next job.
I also took many opportunistic hiking trips to New Hampshire when the weather was particularly nice. I never regretted this, as I got to spend quality time with family and found it refreshing enough that I easily made up the missed time on the next day.
There were also many days when I found myself extremely unproductive, or procrastinating on some specific thing that I was dreading doing. For these days, just getting started was the most important thing, and I often found that spark by asking Claude to do it for me. Still, many other days went by where I actually did manage to procrastinate the whole day. Those days, you just have to accept that there’s something else on your mind, write off that day, and just let your mind chew on whatever it’s chewing on.
To my wife, for giving me the time, space, and encouragement to try a startup.
Fred, for giving me emotional permission to move on from Cartesian Tutor.
Loyal readers of my blog and mailing list, for the many great suggestions, feedback, and connections.