The problem I kept coming back to: a kid can get any answer in ten seconds now, and at the end of it nobody, not the kid, not the parent knows anything more about what they actually understand. So I didn't build a smarter chatbot. We built the boring thing underneath: a real grades 4-8 curriculum, and a concept model where every single question a child answers is tagged to a specific concept. Mastery is scored with a confidence floor, so nothing gets called "learned" off two lucky guesses. Two things changed our approach along the way. The first was the parent report. It would have been easy to hand the week's data to the model and let it write a nice summary. We didn't, the metrics decide everything and the model is only allowed to phrase what they already decided. It cannot state a number we didn't compute. If you're going to tell a parent how their kid is doing, the AI doesn't get to improvise. The second was bigger, and it came from our own data. We watched kids do 100+ interactions in a day and then vanish by week two. Software alone wasn't holding them; the ones who stuck had a person paying attention. So instead of competing with human tutors, we gave them a free portal into the same account — they see which concepts are shaky, assign the next lesson, and send the parent a report under their own name. We take no cut of their rate. Happy to answer anything, including the sceptical questions about whether AI should be teaching kids at all.

Skolvio

1:1 teaching, right at home

1:1 teaching, right at home