Point it at whatever course you are already taking, on whatever screen you are taking it on. It reads the question, remembers the lecture, and puts a validated answer in front of you — as a card you submit, never as a click it makes for you.
A vision model returns typed JSON — screen kind, exact stem, options — and the answer is checked against a rubric before it becomes a card.
It only sees pixels and audio, so a web LMS, a desktop app, a PDF and a TV behave the same.
A persona file you own, lessons learned from your corrections, and a bounded reflection pass that nudges the voice without letting it wander.
Changed frames are classified into a screen kind, a question and its options.
Chunks are transcribed so the lecture becomes evidence a draft can cite.
Slides, transcripts and readings become per-course facts and key terms.
Quiz, test, essay, lecture or reading — routed, drafted, and recorded for review.
The agent decides and drafts. It does not submit — that boundary is deliberate, and it is documented in the README.
Part of the Innotel Platform Stack · docs/stack.md · MIT