What it's like building AI tutors, remote-first
The AI Tutor's five personas — mentor, professor, socratic, storyteller, coach — are built by a small, fully remote team who ship async by default. That's a simple sentence to write and a genuinely different way to work than most product teams default to, so this post is a shorter, more reflective look at what that actually means day to day rather than a highlight reel.
Five personas are really one hard problem, five ways
It's tempting to think of "five personas" as five separate features, but in practice they're five different answers to the same underlying design question: how much should the tutor tell you versus make you work for it? A socratic tutor withholds the answer and asks a better question instead; a coach pushes momentum and encouragement over precision; a professor leans into structured, thorough explanation even when it's slower. Building all five well means resisting the urge to let one persona's voice bleed into the others — a socratic tutor that occasionally just tells you the answer isn't a socratic tutor anymore, it's a professor wearing a costume.
Async by default changes what "shipping" looks like
On a team that treats async communication as the default rather than the exception, a feature doesn't ship because everyone got in a room and agreed it was ready — it ships because the discussion, the review, and the decision all happened in writing, at whatever pace each person's timezone and schedule allowed, and the written trail is the actual source of truth afterward. That has a real cost: decisions take longer to converge than a live meeting would. It also has a real benefit that matters more for a small team: nobody's context depends on having been in a specific room at a specific time, which means the team can be small without any one person becoming a single point of failure for institutional knowledge.
What a normal week actually involves
Less glamorous than it might sound from the outside: a lot of the work is reading real usage patterns from actual tutor sessions, noticing where a persona's tone drifts or where a learner clearly wanted a different kind of help than the persona they picked defaults to, and iterating on the underlying prompting and model routing rather than shipping flashy new surface area. The unglamorous maintenance work — keeping five distinct voices consistent as the underlying models change underneath them — is most of the actual job.
Why remote-first fits this particular kind of work
Building conversational AI systems rewards long, uninterrupted stretches of focus more than it rewards rapid back-and-forth meetings — reading a transcript closely, noticing a subtle tonal inconsistency, and rewriting a prompt to fix it is not work that benefits from being time-boxed into thirty-minute syncs. A remote, async-first structure isn't a perk bolted onto this team; it's a working style that happens to match the actual shape of the problem being solved.