Story 01 · The one that worked
The build I'm proudest of. Coming soon.
I'm writing this one now.
Long-form: the situation, the calls I wasn't sure about, and what shipped. It'll live here when it's ready.
/writing Posts, essays & stories
Posts, essays, and two long-form stories on building with AI.
New essay/stories Long-form
The calls I wasn't sure about, what I'd do again, and the one I walked away from. One is up. The other is coming soon.
Story 01 · The one that worked
Long-form: the situation, the calls I wasn't sure about, and what shipped. It'll live here when it's ready.
Story 02 · The pivot
It started with a problem I could see all around me. AI can very easily hurt a student's success. It's never been easier to delegate to AI when the due date is in 30 minutes and not learn a single thing. I watched it happen in my classes, and honestly, I felt the pull myself.
So I set out to fix it. The idea was to build a harness, the layer of tools, prompts and context that sits around an AI model, that would turn AI into a tutor instead of a delegate. It wouldn't just do the work for you. It would help you actually learn it. I called it Homeroom.
I started by studying the best open-source harnesses out there, like Hermes, OpenClaw and Pi. I pulled them apart to see how each one worked and what I liked about it, and along the way I learned a lot about harness engineering. Then I specced out my own harness, thinking it was going to be revolutionary for student success.
Then I went and talked to the people I was building for. I interviewed students in my cohort, and the problem they most wanted solved wasn't learning. It was something narrower: double-checking a drafted assignment against the rubric and syllabus and projecting the grade they'd get. That was genuinely useful to them, and they wanted a product that could do it.
But two things came up that I couldn't ignore. First, students weren't paying for anything. Who knew that college students' willingness to pay would be so low? I thought about selling to universities and school boards instead and having them roll it out to students. That led to the second, bigger problem. What students most wanted wasn't anything an off-the-shelf harness couldn't already do. The tools they had access to could already check a draft against a rubric. They just didn't know how to set them up to do it.
That made me question the future of Homeroom. I started thinking about what students could build on their own if they knew a little about harnesses, context engineering and what today's models can actually do. The gap wasn't another product. It was knowing how to use the ones they already had.
That's when I made the call. Killing a product is hard. But building the wrong thing is worse, and so is finishing a product and having no end users. That's why I chose to pivot.
At first the pivot pointed toward education: helping students understand AI's capabilities so they'd be better equipped to use the off-the-shelf tools they were already using or that their school provides. As I kept pulling on that thread, I came across the idea for my SI.
SI stands for superintelligence, and the idea behind it is that you should own your own superintelligence, not rent it from someone else. In practice, my SI is an MCP server, a plug-in that connects to the AI harness you already use. It interviews you about your day-to-day work, writes evals for that work behind the scenes, runs all the different models against them and benchmarks the results. That way you can easily weigh cost against performance and use the best model for the tasks you actually do, instead of the benchmaxed models that come out every six hours.
It's built for high-agency people, the ones who don't wait for a tool to decide how they work. They shape their tools around their goals, and I think AI should work for them the same way.
Looking back, Homeroom and my SI come from the same place. I realized that what I wanted to be doing was helping people get better at using AI. My SI is the way to do that.
Stay tuned to see where it goes.