One course. Six weeks. Free. Your own number.

Other AI tools write for a class they have never met. Zelvi has met yours. And it reports back, with numbers, on whether its help stuck.

Why guarded AI

Nearly 1,000 high-school maths students. With plain ChatGPT, practice scores rose 48%. Once it was taken away, exam scores fell 17% below students who never had it. A tutor built to give hints, not answers, largely removed that harm. Bastani et al., PNAS, 2025.

A randomised trial in a Harvard physics course. Students using a carefully designed AI tutor learned significantly more, in less time, than in an active-learning class. Kestin et al., Scientific Reports, 2025.

AI helps when the student still does the thinking. Zelvi is built that way.

A hint comes before any answer
After a first miss the answer stays hidden, so the retry is theirs
The mistake gets a name, taught from your own lesson
Practice is marked by rule, not by the AI

How it works

Student: Gets it wrong. Zelvi names the exact mix-up and re-teaches it with a picture, from your own lesson. Then they try again.
Teacher: Sees who is stuck. One colour map of the class: who has it, who needs help, and the ideas the class is stuck on.
One click: Fixes it tomorrow. An exit ticket aimed at those exact ideas, and a short note for every family. Ready to send.

The offer

Your effort: About 30 minutes to set up: upload one unit and share a join code. No IT ticket, and no LMS integration needed to start.
Agreed before day one: Two or three success measures, for example: share of mistakes fixed on the next try, time from a mistake to its fix, and teacher prep time saved as your teachers report it. Optionally a comparison section, or the same unit test before and after.
Week 2: Your class map: who is stuck, on which ideas, and what to reteach.
Week 6: Your evidence report, built on your own students, with sample sizes shown and ready for your research office.
Cost and risk: Free. When it ends you can export your students’ data or have it erased.

What the report measures

First-retry recovery. Share of struggles resolved on the student’s next attempt.
Eventual recovery. Share resolved within the term, and how long it took.
Repeat misconception. Students who made the same mistake again after help.
Hand-offs to people. How many reached a tutor or instructor, and what happened next.

Safe for schools

Student names are never sent to the AI
Each school’s data kept apart in the database
Your material never trains AI models
Crisis words stop tutoring and alert staff