One read under every surface.
Everything ClassGrade does runs on a single operation: locate a point in the learned geometry of English, then read the true neighbourhood around it. The live demo shows that read directly — the skills on either side of a point, retrieved, then written up. What follows is what that one read makes possible.
See the live demo →One read, many products.
Everything you just watched is a single operation: locate a point, retrieve the true neighbourhood around it. That one read is the engine under everything ClassGrade does.
Tells a learner what to work on next. The forward path, read as guidance: what to focus on now, and which gaps behind are holding it up.
Grades work against outcomes. The same read resolves a piece of student work to its position — the marking you saw in ExamBuilder.
Sequences content into a path. Walked end to end, the neighbourhood becomes an ordered route through a subject — outcome, fluency, and sequence in the right order. A book, as data.
Generates the material at each step. Content produced to fit each position, grounded in the retrieved neighbourhood — the same retrieve-then-write you saw above, turned to authoring.
The retrieval is the same every time. Only what you ask it to produce changes.
It runs at the edge.
The same read that powers everything above compiles to an artifact small enough to run on-device, against a local model — no cloud, no round-trip, no backend to depend on.
For emerging markets, that means guidance and grading work where connectivity doesn’t: the engine sits on the device, not behind a network it can’t rely on.
For sovereignty and ESG deployments, it means the strongest possible version of partner-controlled data — the artifact runs in-country, on the partner’s hardware. There isn’t a data path to secure, because there isn’t one. The partner holds everything.
It’s not a roadmap line. The artifact is real and it runs locally today — we’ll show you.
Why voice works here when it doesn't elsewhere.
Most voice tutors are weak at the one thing that matters: reasoning about where the learner actually is. Every turn, they hand the transcript to a large model and ask it to work that out from nothing — slow, costly, and often wrong.
Here the activity already knows its position. The transcript and its phonemic read resolve straight to a coordinate, and that coordinate is the context — no search, no retrieval step, no large-model round-trip to figure out the learner from scratch. The geometry has already done the reasoning.
That's what makes it run small and local. The hard part — knowing where the child sits — is solved before the model speaks, so a compact on-device voice model is enough. Low latency, no cloud dependency, and it holds up: tested locally on a 1.5GB voice model, it works.
Underneath every surface above is one framework.
BlueberryML maintains a Riemannian-manifold model of learning with a five-dimension coordinate system for English assessment, and field-theoretic quantities that drive adaptive logic. Every capability on this page reads from and writes to the same substrate.
The mathematical framing and the company posture sit on blueberryml.com. This site stays on what the framework does in deployment.
Read about BlueberryML