How We Replaced a Queue of Mentors Answering Doubts With an AI Tutor That Never Sleeps

We drew it six different ways. Mentor front and centre. Mentor tucked to the side while a chat did the talking. A full grid of guides to pick from, like a classroom you could scroll through. Each layout was us asking the same question a little differently: what does help actually look like on a screen?

It was warm, and it was unmistakably human. For a first bet, that mattered — a face is easy to trust, and trust is the whole game when a student is about to admit they are stuck.

But watching real learners use it taught us something we could not have designed our way to. They did not want to leave their work to go and have a chat with a mentor in a separate window.

They wanted help on the exact line that was fighting them — without losing their place, their code, or their train of thought. The help was good. The place was wrong.

So the face, lovely as it was, turned out to be the least important part. What people wanted was proximity: a helper that was already there, not a destination they had to travel to.

That single insight reframed the entire product. The next version would not ask students to come to the tutor. The tutor would go to them.

And there were two effortless ways in. You could click to ask, or you could simply say “Hey Rahul!” out loud and start talking, hands still resting on the keyboard. Asking for help stopped being an interruption and started feeling like turning to the person sitting next to you.

That one move — from a place you visit to a presence that is already there — is the whole story in miniature. The same good idea, shifted six inches closer to the problem, quietly became something students reached for by reflex.

Under the surface, the team treated quality like an engineering problem. Every answer could be rated with a simple thumbs up or down, so the tutor kept getting sharper. Response times were watched to the millisecond, and the prompts and evaluations were tuned until replies came back fast and stayed on point.

This was not a guess. Reading real conversations showed that roughly seven in ten would directly benefit from acting on a single highlighted phrase — the exact moment a learner is most engaged. So we removed the friction right there, turning “retype and re-explain” into a single tap.

The loyalty is what really counts. Around 78% of students came back for a second conversation, nearly half returned five times or more, and better than one in four crossed ten. One especially determined learner had 672. When a tool becomes a habit, it has stopped being a novelty.

Word spread fast, and the first months brought a burst of early fans. Then, instead of coasting, the team did the harder thing: they spent the middle stretch listening closely to how students actually used it.

That listening became the redesign you just read about — the move into the editor, the voice, the conversation states, the act-on-any-line actions. It was less about adding features and more about closing the distance between a stuck learner and a clear answer.

When that work landed, the line did not just recover. It climbed past everything that came before it, reaching 11,809 students in a single month — the busiest the tutor has ever been.

And it settled into two homes, exactly where we hoped. Most of the help now happens right inside coding practice, where close to 29,000 students have leaned on it while solving problems, with a second version living inside full learning sessions for another 14,000. Two places, one job: a patient explainer that never runs out of hours.

That is the real arc of this story. We started by giving help a face, and ended by giving it a place — right beside the work, at any hour, for anyone who asks. The face is gone now. The mentor stayed.