We've built something narrow: an AI that grades TOEFL speaking and writing, not another question bank. One that reads the answer itself, marks every sentence that cost points, and explains each one. Speaking and writing only. Scored on the ETS band. Never a template.
AI-Graded TOEFL Speaking and Writing
A response goes in and the report comes back in minutes. Nothing to book, nobody to wait on, no two or three days between the answer and the reading of it.
A band score says where a response landed. It does not say what to change. Every report here also marks the sentences that cost points and explains each one.
Most people here are preparing after work, with nobody around to read what they wrote. The whole loop runs without a tutor in it.
Every decision we made is rooted in the belief that practice and the feedback should be accountable, actionable, and comprehensive.
Every note we generate is matched back to the student's own submission. On an essay, anything our AI cannot find verbatim is dropped rather than shown; on a recording, the passage a note points at is re-cut from the transcript itself.
Diagnosing a problem is only half of the work. Each issue arrives with a method that addresses it and a way to practice that method immediately, so a weakness is something worked on rather than something recorded.
Preparation should not be limited by what a candidate can get hold of. Practice here spans the situations the test actually asks about rather than one archive of past papers. Our AI sorts out the latest scenarios from social feeds and covers as many categories as possible.
TOEFLAIR has been one person's project since December 2025, rebuilt in January for the 2026 test format. Every number here is a straight count taken on 1 September 2026.
People who have signed up and actually practiced here, not just opened an account.
Recorded and written answers people have handed in, with test and founder accounts left out.
Each report covers one completed practice, which usually holds several answers.
Live practice sets across the four task types, 39 of them built from real exam recordings.
I took the TOEFL 7 times.
The first was at the end of 2018, when I was a third-year undergraduate with no idea about my future. So I picked the option furthest from where I was standing: a master's degree abroad. My English had always been mediocre, if not poor. I bought TPO sets and ground through all four sections, and scored only 89 out of 120.
The application season was closing in, and a higher score sooner meant a stronger application. So in 2019 I paid a Chinese prep service for speaking and writing. I submitted a recording or a piece of writing through their dashboard, and two or three days later it came back from a real tutor, marked up and scored. With that feedback and their material, the course took me from 89 to 98, and its momentum carried me to 99. Then I sat it again without changing anything and came back with a 98.
Generally, 100 out of 120 is the bar people talk about for a Chinese applicant at a top-30 school, and the ones I most wanted were exactly that kind of place: CMU, UCLA, schools with a real research culture around them. UCSD, on my list from the start, did make me an offer. Then the pandemic shut the consulate, there was no way to get in for a visa, and I took a job in China instead. The work wore me down, and it left me wanting to study abroad more, not less. So I kept sitting the test.
But by then there was nobody. What I had was what had worked in the last few attempts: the feedback I'd been given, the templates, a few things I'd worked out myself. I naively assumed that if I kept practicing using them the score would keep going up. I drilled the same fifty-odd TPO sets over and over, judging a session by whether I got the question right and how many sets I'd finished. Neither of those tells you anything about your English. The next two totals fell to 96, and then to 94, and I started to think it was going to stay that way: a job with no room to grow, and no way out of the country.
I went looking for where it had gone wrong. Listening had always been weak, and writing was sliding: 28 on the attempt where I scored 99 overall, then 25, then 22. That 28 had convinced me I could write, when what I actually had was a set of rules for writing. I went back through every practice response I still had, and I could see that none of them had full marks. I could not say what was wrong with any of them.
I was the only one judging my own work, and I was the one person who could not see what was wrong with it.
Reading and Listening have a standard, single right answer. Without exam experience of their own, though, test-takers cannot judge their speaking and writing or know how to improve their performance. What a prep service comes down to is a return to the atomic skills of English, like how to organize your words to support your point. In 2022 I chose a paid prep service again, and put my arrogance aside. Working through their routine, I went back to the basics, refining a single piece of writing until it could earn 28+, and learning to take in what a passage said without translating it into Chinese in my head. In the end that took me from 94 to 106, and helped me get an offer from NUS.
I'm not a teacher and have never taught a language, but AI can do this better. Since ChatGPT came out, models have kept getting better at understanding language and at following instructions. I believe AI can become a perfect partner for TOEFL test-takers, the way the human tutor I had was, giving detailed and objective feedback. That is why I believe AI can solve the problem most test-takers run into.
