Finds the cause, not just the score
Reads the working, gives partial credit, and traces each wrong answer to the misconception behind it.
AI Grader marks handwritten papers from a phone photo and turns them into feedback, practice, lesson plans and school reports.
How it works
Reads the working, gives partial credit, and traces each wrong answer to the misconception behind it.
Every answer counts towards a skill in your curriculum, so progress builds paper after paper, not one test at a time.
Practice sheets print with each student's name. Students who were absent get a catch-up pack.
A London school, an education company in 60 countries, and the Middle East's biggest exam-prep service.
A DfE-registered London school, piloting AI Grader in GCSE and A Level.
Works with 4,000+ schools in 60 countries. Pilots start in January in UK and Rwandan state schools.
The Middle East's biggest exam-prep service. One million of its marked exam papers now sit behind AI Grader.
What that gives youWatch one real class, from the first photo to the printed practice.
And every student sees why they got each mark.
One page or the whole stack. It reads the names too.
Everything else is marked for you. You always have the final say.
One strength, one priority with a practice line, and what to work on next.
The shared mistakes, the groups and the practice are ready before you are.
Start with the shared mistake, group the rest, check in with three students.
Ten questions, about 12 minutes, ready to print.
Every skill, month by month, against the student's own past.
Every paper your school marks is kept and read. Ask in plain words, and get an answer, the papers behind it, and something ready to use.
Answers come from your marked papers, not the internet.
Every answer names the papers it used, so you can open them and check.
Sheets, notes and reports are drafted. A teacher sends them.
Heads of year, school leaders and families each get their own view.
One short note a week: one win and one simple thing to try.
Only after the same thing shows up twice and a teacher approves it.
One page. No rankings.
Ghalamchi is the Middle East's biggest exam-prep service. Every paper is marked by two examiners, and by a third when they disagree. That record of how real examiners mark now sits behind AI Grader.
We measure AI Grader against how often two examiners agree with each other, one question type at a time.
Crossed-out working, cramped margins, phone photos, and right-to-left script mixed with maths.
The wrong answers that keep coming back power every diagnosis and every practice sheet.
Results are read against how students usually do on that kind of question. Faulty questions are flagged before they count against anyone.
It knows which mistakes fade on their own and which stay unless someone steps in, so those come first.
Students keep writing on paper. No tablets, no new systems to learn.
Designed by a research team from LSE, Oxford and the University of Tehran.

Every feature follows the evidence. Four examples:
Students given AI answers on demand did worse in exams. So the full solution opens only after the student has tried.
Bastani et al., 2025Mixed problems beat blocks of the same problem. Every practice sheet mixes in earlier topics.
Rohrer et al., 2020Quick reviews at one, three, eight and sixteen weeks slow down forgetting.
Cepeda 2008; Yang 2021Good feedback adds about six months of progress, if students read it. So the mark comes last.
Butler 1988; EEFAI Grader is made by LiveBook for Schools, with its headquarters in the UK.
Endorsed as innovative, viable and scalable.
Award winner at the London School of Economics' venture programme.

One of the top 20 startups from London's leading universities.
Recognised by the international EdTech forum.
Selected among 2024's education innovations.
No. Marking happens automatically. You only look at the answers it flags, and nothing goes to students or families until you approve it.
No. Every report starts with three things to do at most. The charts are there if you want them.
An unreadable answer is never marked zero. It's flagged for you. Blank questions at the end of a paper aren't counted as “didn't know”, because time may have run out.
No. Groups have neutral names like Cedar and Pine. Each forms around one shared mistake and ends when it's fixed.
English, Arabic and Farsi handwriting, including right-to-left scripts mixed with maths.
A phone, a photocopier and the tests you already set. Pick one subject and one year group.
General models are good at answering one question at a time. Marking a whole school's papers, every day, needs something else.
| ✕General chat model | ✓AI Grader | |
|---|---|---|
| Same answer, same mark | A second run can give a different mark and different feedback | One scheme, applied the same way across hundreds of papers |
| Checked against examiners | Not measured against how your exam is marked | Measured against how often two examiners agree |
| A history of every student | Remembers chats, not skills | A skill map for every student and the whole school |
| Cost at school scale | Climbs fast with hundreds of papers a day | Built for daily use at a low cost per paper |
| Right-to-left handwriting | Often misreads Arabic and Farsi digits | Built and tested on it, mixed with maths |
One subject, the papers you already set, and no set-up cost.
Use the tests you already set.
You only look at the unclear answers.
Which gaps we targeted, and how many closed.
Each one comes with a line you can share in a staff meeting.