loading...
Graide's AI grades university exams and essays with 99% alignment to human instructors, giving educators quality feedback at any class size.
ENTRY ANGLES
AI-powered feedback system that learns instructor-specific teaching patterns and scales personalized feedback · Interactive video lessons with embedded comprehension checks and adaptive error responses · Platform that categorizes typical student mistakes by topic to minimize instructor time per feedback
VERTICALS
CAPABILITIES
AI/ML models trained on instructor-specific feedback patterns and teaching approaches, Video platform with interactive lesson capabilities and branching logic, Error classification and taxonomy system for topic-specific common mistakes
Graide is an AI-powered platform that helps educators grade student assignments and exams.
The platform is built for universities, examination boards, and professional certification programs.
Its AI engine handles math, physics, biology, chemistry, economics, and engineering disciplines.
Recently, the startup added essay grading. In testing, the AI's marks aligned with human instructors 99% of the time.
Graide doesn't just score work – it provides feedback, flagging where errors occurred and suggesting cleaner or more conventional approaches. The platform can also read handwritten mathematical notation, which matters a lot for grading physics and math assignments where students work through equations by hand.
The startup was founded in 2021 by former students who, during their final years of study, worked as teaching assistants doing the tedious work of checking homework and grading tests on behalf of professors. Worn down by the manual effort, they set out to automate it.
Graide is currently used by several UK universities. The company has recently run its first pilots with US institutions.
Graide has now raised £1.7M, bringing total funding to £2.5M.
In principle, you could build a naive version of this with a general-purpose AI model – upload the assignment, prompt it to grade and comment. But Graide works differently.
The platform's AI looks at the student's work, understands it, and then searches a database of previous feedback given by that instructor on related assignments:
- If the same conceptual response has appeared before and the instructor has already given feedback on it – the AI drafts its response based on that existing feedback automatically.
- If the work resembles previous submissions but isn't identical – the AI generates a partial response based on the matching portions and hands a draft to the instructor to complete the unfamiliar sections.
- If the work is entirely novel – the AI does nothing and passes it to the instructor to handle from scratch. That feedback then enters the database and informs future grading.
The threshold for what counts as "similar enough" for automatic grading is configurable by the instructor, so they control how much the AI handles independently.
The result is a system that learns with every manually reviewed assignment – and eventually handles the overwhelming majority of submissions automatically.
The clever angle: teaching is inherently cyclical. Year after year, instructors cover the same material. And year after year, students make the same mistakes. Graide groups the incoming stream of assignments by similarity – which saves time even before any AI grading kicks in. Once the same errors repeat often enough, the AI can handle them without any human involvement.
The impact shows up in the numbers. Average grading time per assignment drops from 11.2 minutes to 2.8 minutes. Average feedback length increases from 23 words to 166 words – because the AI never gets tired of writing the same explanation again, while human instructors naturally start cutting corners when they've typed the same thing for the hundredth time. Students end up receiving more substantive responses that actually help them understand their mistakes.
Feedback is the most important component of learning. Without it, anyone could learn anything by watching tutorials online.
But feedback is precisely where instructors spend most of their time. More feedback means better learning outcomes – but it also means more hours, which can make the economics of teaching unworkable.
Instructors face an eternal tradeoff: give thorough feedback and work at a loss, or scale back and let students who need the most help fall through the cracks.
General-purpose AI doesn't fully solve this, because good teaching is idiosyncratic. Weak instructors just rehash textbooks and accept textbook answers – something a generic model trained on those same textbooks can easily evaluate. But strong instructors have their own approach, and their feedback reflects that.
The insight about typicality of errors, though, is a genuinely powerful one. The number of typical mistakes students make on a given topic depends on the topic itself, not on the number of students taking the course. That means a course could scale from 10 students to 1,000 with almost no increase in instructor time spent on feedback.
Kyron Learning, [covered here](/review/teper-vygodnee-ispravljat-chem-objasnjat) last November, applies the same principle to video instruction and raised $20.1M. Their platform delivers interactive video lessons: after each concept is explained, the student answers comprehension questions. Correct answers advance the lesson. Errors trigger a short video that specifically addresses the mistake. Experienced instructors, who already know the common misconceptions, pre-record those correction clips in advance – so the platform assembles a personalized lesson path for each student on the fly.
Education is a permanent market, and feedback is its most critical – and most time-intensive – component. The opportunity: platforms that automate feedback by exploiting the predictable, repeatable nature of student errors.
Graide and Kyron Learning offer two models worth examining – either to replicate directly or to adapt into a different subject area or learning context. What other domains could benefit from the same approach?