ORBIT

The problem

I wanted AI help with studying and planning that actually remembers my courses, deadlines and weak topics, without paying per-request API costs. Off-the-shelf chat apps forget everything between conversations and can’t touch my calendar.

What I built

ORBIT is a hub of four AI agents (Study Tutor, Planner, Daily Brief and Content) that runs as a single page on claude.ai, using my existing Claude plan instead of a paid API. The agents share a memory of notes, weak topics, deadlines and preferences, stored in the page’s own database.

• Calendar: ORBIT reads the next two weeks from Google Calendar. Agents can propose new events or removals, but nothing changes until I tick each one on a confirmation card.
• Course material: I can upload PDFs, notes or photos of slides. The tutor gets whole documents when they fit, or the best-matching passages (BM25 search) when they don’t, and it cites page numbers in its answers.

What I learned

• Performance matters even in a dashboard: the first version made my laptop hot. A single frame governor with per-layer frame-rate caps cut idle main-thread time from 291 to 90 ms per second.
• AI actions need a human in the loop: agents only ever propose calendar changes, and the page checks every event id against my real calendar before acting.
• Working inside hard limits: prompts are capped by size, so conversation history is trimmed by real byte count, and long documents fall back to excerpts.