Starship
DropOut

I shipped an app without knowing how to code. It never met a user.

Six months building a mentor that takes the initiative, five months of work never deployed, and a database deleted before the first measurement.

I shipped a complete application to production without ever having written a line of code before. It stayed online for five months. Not a single user signed up, and its database was deleted before the first measurement was ever possible.

Here is what it was, and what its failure taught me — which is not what I thought it would be.

Every study tool makes the same mistake

AI-assisted study tools share an implicit architecture: the student asks a question, the tool answers. That architecture assumes a student who already knows what they don't know.

But a student who is falling behind stops asking questions. They don't know what to ask. The most capable model in the world stays silent in front of a silent user — and that is precisely the user who needs help.

DropOut came out of that observation. The product was meant to invert the initiative: a mentor, named Ben, who introduces an idea before it is asked for. Create curiosity instead of answering it.

It's a thesis about the interface, not about the model. The bet fits in one sentence: the difference is decided by who speaks first.

What was built

The mentor exists in several copies. Up to six distinct Bens, one per course, each with the context of its own material. Above them, a cross-cutting Ben General, who sees every course, exam and assignment.

That distinction matters more than it looks. Answering "what should I work on tonight?" requires seeing the calendar, not just the chapter. A per-subject mentor cannot arbitrate between subjects; someone had to sit above them.

The student uploads their course PDFs, and Ben reads and quotes them exactly, instead of paraphrasing from memory. Around that: automatically generated quizzes, an exam countdown, checkable assignments, push notifications, an animated onboarding, and installation on iPhone and iPad.

The stack: Next.js 16, TypeScript, Tailwind, Supabase, the Anthropic API, Vercel for hosting, OneSignal for notifications.

One choice is worth calling out: two models, split by task. Sonnet for conversation, Haiku for quizzes. Conversation justifies the most capable model; quiz generation is a constrained task where the fast model is enough and costs a fraction of the price. That was the decision that made the product economically viable.

A Safari bug was fixed with a combination of API timeouts, error boundaries and progressive message loading — three fixes treating the same symptom at three different levels.

The two things I refused to add

Along the way I ruled out two features: consecutive-day streaks, and a "Survival" mode.

These are engagement mechanics. They reliably increase usage metrics, and that is exactly the problem: they increase them without proving anything about learning.

A student who opens the app so as not to lose a forty-day streak isn't learning. They're maintaining a counter.

Refusing those mechanics is consistent with the product's thesis. It is also a decision I never put in front of a real cohort, so I have no proof it was right. I made it on conviction, not on measurement — which is precisely the charge I level at everyone else.

What happened next

The product shipped. Then the project stopped without an explicit decision — and this is where the dates become instructive.

MeasureValue
Last production deploy19 April 2026
Last local development24 September 2026
Work never deployedfive months
Users measurednone

Five months of fixes and improvements that never reached the live version. I kept building for a product I had stopped updating.

Meanwhile, the Supabase project serving as the database was deleted. The address the application called no longer resolved. The landing page rendered, the sign-up screen too, but no account could be created.

The application was online and out of service. The deployment has since been taken down.

What I never tested

This is the most important section of the article, and it is entirely empty.

  • The thesis itself. Does a mentor who takes the initiative hold a struggling student better than a tool that waits for the question? No data.
  • Retention. How many students come back after a week, after an exam session.
  • The effect of my refusals. Without streaks or Survival mode, does usage hold up over time, or did I deprive myself of the thing that brings people back?

All three questions are out of reach today without rebuilding the database. The count of users who might have signed up disappeared with it.

What I take from it

The rare skill isn't having the idea, it's shipping it. DropOut proves I can go from idea to deployed product without knowing how to code at the start. That's something I didn't know about myself in April.

The corollary is less comfortable, and it's the real lesson.

Shipping is not validating. A product in production with no measured user is still a hypothesis, better dressed. And a hypothesis you leave to sleep ends up losing its infrastructure before it ever gets an answer.

I spent five months perfecting something nobody was using, instead of spending one week finding twenty people who would use it. It is the most ordinary mistake there is, and I made it in full.

To pick it back up

  • Step zero: recreate a Supabase project and apply the schema. Nothing is possible before that, and it will mean restarting from the branch frozen in September, which holds five months of fixes that never went live.
  • Launch planned the way Facebook launched at Harvard: one campus first, where I can recruit in person.
  • Then the only step that counts: twenty real students during an exam session, and measure who comes back. Fix the success threshold in advance, before looking at the numbers.

One project a week, in your inbox

What was built, the numbers it produced, and what did not work. No other email, and one click to leave.