Eedi shares real student learning data with researchers worldwide — and does it without exposing a single student. Here's how pseudonymisation and our PIIvot framework let us open up genuine tutoring conversations, and why locking that data away would cost students more.
Most assessment stops at right or wrong. By integrating our Diagnostic Engine into Imagine IM's Knowledge Checks teachers get insights into the specific misconception behind each error a student makes, and accurate predictions about the misconceptions in the way of what students learn next.
Case Study
Eedi's Misconception Graph is now available as an open-source public good through Learning Commons. As of today, a mapping of more than 8,000 common maths misconceptions, informed by 200 million real student responses, is free for anyone to build on, under a CC BY 4.0 licence.
News
Anyone who has worked with children knows that their voices matter. Not merely as a 'tick box exercise', either, but what they have to say actually makes a difference. It guides how you teach a concept to a class or how you organise a room. Children’s thoughts and opinions impact how you communicate with them and what you choose to focus on. They can change your mind.
Blog
Research
At Eedi, we only scale what works. This exploratory study in Rajasthan builds on RCTs focused on diagnosing and resolving misconceptions over the last 5 years that provided gold-standard evidence of efficacy in UK classrooms.
Insights into our participation, co-hosted events and favourite papers from this year's Festival of Learning in June, in Seoul, Korea.
Claude for Teachers is the newest place for teachers to find easy access to Eedi’s Diagnostic Questions.
Through a Model Context Protocol (MCP) integration, teachers can now access Eedi's full library of diagnostic questions - representing a decade of research into maths misconceptions - directly within Claude.
Press Release
This trial ran over two academic years and leveraged our testbed platform, Eedi School, as a classroom and homework tool. By 12 months the effect was on the cusp of significance, it crossed into clear statistical significance at 18 months (d = 0.46) and held at 24 months (d = 0.30).
This blog summarises newly published research from Eedi on selective prediction for responsible knowledge tracing. It discusses how model uncertainty catches risky predictions, preventing silent failures and routing teacher effort to critical intervention points.
This blog summarises our new research paper which explores two novel approaches for extracting latent knowledge state from student-tutor dialogues, leading to AI tutors that are less verbose and more effective at guiding the dialogue.
Eedi is a grantee of Accelerate's '26-27 Call for Effective Technology (CET) program to bring an RCT of our constrained AI Tutor into US classrooms for the very first time.
This blog announces our second "constrained" AI Tutor RCT, started in April 2026, in partnership with Google DeepMind's research team. At its heart is Eedi's diagnostic engine which powers the testbed tool being used in classrooms, Eedi School.
Two invitation-only gatherings - One at Stanford University, the other in Kigali, Rwanda - brought together researchers, policy makers, educators, learners, and entrepreneurs to discuss AI in education.
We believe teachers have a distinct advantage when it comes to using AI to build effective teaching resources and that their expert human-in-the-loop role is key from a safety and pedagogy perspective. Chief Academic Officer, Craig Barton, explains why.
Pedagogy
Newsletter
How specialised Knowledge Tracing (KT) models continue to outperform general-purpose LLMs, in terms of accuracy, speed and cost, when used in learning tools in real world classrooms.
.png)
