Improving Cereby Capabilities: From Plain Text to Rich Visual Learning Materials
How we aligned AI-generated study content with the same rich surfaces users already expect from professional notes and quizzes.
The problem was the medium, not the model
Cereby AI could already reason about learning context. What landed on the page was the problem. Flat bullets. Math rendered as Unicode hacks. No diagrams. Layouts that felt cropped or amateur. The ideas were often sound; the presentation undermined trust and killed reread value.
The instinct was to keep tuning the model. The right move was to fix the output layer first.
The shape we landed on
We built in three phases, each one feeding into the next.
The key decision was not to special-case AI output. When Cereby uses the same editor, templates, and renderers as learners, consistency compounds. Every improvement to user authoring automatically improves generated notes, quizzes, and flashcards.
What we built
Real math rendering was the step-change. Inline and block LaTeX with broad symbol coverage replaced the Unicode approximations that made STEM content embarrassing.
Structured templates encoded the pedagogy directly: Cornell notes, concept maps, comparison tables, and timelines, with block-level primitives for definitions, proofs, code, and callouts. A table of contents and section progress navigation came for free.
The rest of the surface area (a subject-aware visual design system, diagrams, charts, a WYSIWYG equation editor, and rich quiz and flashcard types) filled out the stack so users and AI share the same authoring affordances end to end.
What changed
| Signal | Before (approx.) | After (approx.) |
|---|---|---|
| Note quality (user rating) | ~2.5 / 5 | 4.5+ / 5 |
| "Visually appealing" | ~30% agree | 90%+ |
| "Easy to read" | ~60% | 95%+ |
| Formula render correctness | ~40% | ~100% |
| Note creation rate | baseline | +50% |
| Revisit rate | baseline | +70% |
What this taught us
Build for humans first, then hand the same tools to AI. You get two rounds of quality improvement from a single investment, and the consistency compounds.
Templates encode pedagogy in a way free-form generation cannot. Cornell notes and concept maps outperformed prose dumps on revisit rate regardless of underlying content quality. That is not obvious from the model side.
Visuals reduce abstraction cost in ways that do not show up until you measure them. Charts and diagrams were among the highest-leverage additions, especially for concepts that do not compress well into text.