AI Canvas
A typed AI workspace where a chat assistant extracts schema-typed structured objects from conversation into a queryable store — with row- and field-level provenance, so you always know whether a value came from you or the model.
React RouterDrizzle ORMPGlitePostgreSQLAnthropic APITailwindshadcn/ui
AI Canvas pairs a chat loop that replies in typed blocks — prose, extracted object sets, schema proposals, query results — with a workspace of structured objects extracted from the conversation.
The interesting parts are in the data model:
- Provenance everywhere. Every row and field records where it came from. Inline corrections are tracked, and user writes outrank model writes.
- A versioned schema registry. Schemas live in an append-only, versioned JSON-Schema registry, so extraction types can evolve without breaking existing data.
- Queries you can see. Natural-language questions compile to a visible, editable filter IR rather than opaque magic.
It also runs zero-config locally on in-memory PGlite and produces deterministic mock output without an API key, which makes development and testing pleasant.