Keyi: Aha-First Multimodal Courses from Any Document
A personal project outside campus: a browser-based agent that turns teaching material into readings, 2D interactives, quizzes and 3D scenes built around the few insights a learner must walk away with.
2026, a personal project outside campus
Background: courses that look great and teach nothing
Generative AI has made it effortless to turn a document into something that looks like a course, but not into something that teaches. Modality choices follow whatever is cheap to generate, nothing guarantees that the material covers the intended learning goals, and learners are left to wander through the output.
I hit this wall myself. Keyi began as a no-code VR and 3D teaching builder for teachers, built on the belief that fancier and more is better. Teachers did not say they didn't want it; they said they could not fit it into a lesson and could not tell what their students had learned from it. Once the visual novelty wore off, the scenes taught nothing.
Good is not matched, and more is not matched.
How Keyi works
Keyi first distills the small set of "aha" insights a learner must leave with and anchors them in a machine-checkable knowledge structure. Only then does it produce content, and every section is accountable to the goals it claims to teach. 3D is one medium among several, chosen where it helps, next to readings, 2D interactives and quizzes. A learning companion then guides learners through the result step by step.
Under the hood, a planning and execution agent architecture with 20 dynamically loaded skills and 40 tools keeps generation reliable, and it shares its teaching patterns and action vocabulary with the
AI PBL Agent.
Study
We evaluated the generated learning experiences in a remote usability study across five subject packs with six participants, using pre and post quizzes, workload and satisfaction questionnaires, and interviews. Mean quiz totals rose from 1.5 to 23.8 out of 30 and satisfaction averaged 6.4 out of 7. Participants named the interactive and 3D sections plus the quizzes as what made them learn; thin reading scaffolds and occasional 3D fidelity issues were the clearest frictions.
My Major Roles & Contributions
- Led the project from the failed VR builder to the current aha-first design.
- Designed the planning pipeline, knowledge structure, and how pedagogy constrains generation.
- Ran the usability study and wrote up what it supports and what it challenges.
×