AI Open Studio
A collaborative learning and research workspace combining a Mistral prompt engine, structured document AI, shared Yjs state, and native peer media.
The operating problem
Learning and research tools needed shared state, safe document changes, and live media without disguising single model calls as agents.
My contribution
Built 30 prompts across 10 categories, constrained BlockNote operations, research rooms, a custom Yjs transport with 50 ms batching, reactions, drawings, and native WebRTC signalling.
System shape
The key decision
Document edits are restricted to append, prepend, or rewrite; destructive rewrites require confirmation and preserve undo history.
Reliability by design
Yjs updates are batched and merged before broadcast, full state reaches joiners, and awareness is relayed separately.
What the source proves
Honest limits
Normal documents remain last-writer-wins. The peer media system has no TURN or SFU claim and is not presented as large-room infrastructure.