Taxbox field notes
Working notes from shipping a multi-tenant document engine — isolation as a product constraint, not a filter bolted on later.
Context
Tax and compliance teams drown in unstructured PDFs. Manual classification is slow and hard to audit. One wrong multi-tenant join is an incident — so the stack has to make cross-tenant access structurally impossible.
What we tried
- JuiceFS for large document blobs with predictable IO.
- Tenant-scoped metadata and vectors on Turso/SQLite.
- Docling + model path for ingest and classification.
- Desktop pairing so firms keep vault workflows on machines they control.
What stuck
Isolation is a product feature. Operator UX stays calm under deadline pressure — status chips and hierarchy over terminal fireworks.