The tutor that doesn't exist in most classrooms.
UNESCO counts more than 250 million children who don't meet basic math proficiency. Many of them sit in classrooms with sixty other students and one teacher, in places where the internet is something that visits, not something that stays. Off-the-shelf tutoring apps assume broadband, an iPad per student, and a parent in the room. The kids who need a tutor most are the ones those apps were never built for.
EduScribe AI is built for those classrooms. It runs on the device the school already owns, listens through the noise, reads handwritten work, and never asks for a Wi-Fi password.
Your worksheet. Read.
A student photographs the page they're stuck on — handwritten or printed, neat or scratched out, on paper or on a whiteboard. EduScribe reads the equation, the student's working, the answer they wrote, and the place where it went wrong. Not just the wrong answer; the wrong step.
Your tutor. Talking.
Then the agent asks a question. It does not give the answer. It does not explain the whole concept from scratch. It asks the one specific question that gets the student to see what they missed: What happens if you add the numerators only? What does the denominator represent? The student speaks back. The agent listens. The conversation continues until the student gets there themselves.
And when the agent reads math aloud, it pronounces it the way a human teacher would — "three-fourths," not "three slash four." π is "pi," not "pi symbol." The student learns the right name for the thing the first time.
Your classroom. Connected to nothing.
The whole loop runs on the device. There is no API key, no latency, no "the tutor is offline, please try again later." A school with one Mac in the corner of the classroom can serve sixty students all day, on a Tuesday in a town with intermittent power, and the experience is the same as one in a top-tier urban school with gigabit fiber.
The pedagogical promise
The tutor never solves the problem for the student. It cannot. The way we built the agent literally does not let it produce the next correct step until the student has tried one themselves. A tutor that gives away the answer is worse than no tutor at all — and the kids who need a tutor most don't deserve a worse one.
What that means in practice
- No giveaway answers. The agent identifies the misconception and asks the right question. The student does the thinking.
- No internet required. The school's broadband can drop, fail, or never have existed.
- No data collection. The student's photos and conversations stay on the device.
- No vendor lock-in for the school. Open the app, use it, close it. There is no account, no quota, no per-seat fee.
By the numbers
EduScribe sustains 120 tokens per second on current M-series hardware, which is fast enough that the student never feels the model thinking. The tutor reads handwriting, listens to spoken questions, and replies aloud — three input modalities in a single tutoring loop. Zero cloud calls means a school can run it all day with no recurring infrastructure cost.
Want one of your own?
EduScribe is the case for treating an LLM as a deployment constraint, not a capability constraint. The interesting design work is not in picking the biggest model — it is in shaping the agent so it can't do the wrong thing for the user. If your product has a behavioral contract — a tutor that should never solve, a triage agent that should never diagnose, a coach that should never advise — the same patterns apply.