Frontier intelligence for Education.
Agentic education at scale needs the best data. Evals, verifiers, datasets and learning environments that push AI tutoring and long-horizon tasks in education systems.
Four products to push Agentic Education.
One pedagogical engine for tutors that teach, and for agents that run long-horizon tasks inside education systems.
Find where the tutor and education system fails.
Evals
Measure how well a model actually teaches, turn by turn and across weeks of simulated learners. Graded by expert educators and verifiers, with a failure report that shows where the tutor breaks.
- Multi-turn evals
- Human eval
- Learning gains
Rubrics & verifiers
Rubric-based graders that score every tutor turn and check real mastery and task completion. The same signal becomes the reward when you train.
- Rubrics
- RLVR
- Mastery checks
Close the gaps evals surface.
Learning environments
Simulated students and education systems where agents practice long-horizon tasks: planning a course, tracking a learner for weeks, closing the loop.
- RL environments
- Simulated students
- Long horizon
Datasets
Expert-built SFT and preference data over full learning trajectories, aimed at the gaps our evals surface.
- SFT
- RLHF / DPO
- Learning trajectories
Off-the-shelf or co-designed with your research team. Public benchmarks built with top universities.
We built our own AI tutor and run it with live learners. Our ground truth comes from real students, not synthetic ones.
We are tackling pedagogy slop first with the labs, then as an API for everyone else.
Measuring whether students actually learn.
Public benchmarks, built with academic partners.
Global Education Benchmark
In developmentHow well do frontier models teach across countries, languages and curricula? Not what they know, but whether students learn.
In partnership with
AI Tutoring Benchmark
In developmentPedagogical moves, learning gains and long-horizon support: a benchmark built to measure tutors, not test-takers.
In partnership with
Let’s end pedagogy slop.
Tell us where your model or tutor falls short. We come back with an eval, a pilot scope and a timeline.