From a folder of files
to a question you can ask.

Drop CSV, Excel or zip files — several at once is fine. Grovely reads the structure: encoding, delimiter, header, row count, and a type for every column from its first 10,000 rows. Then it shows you the plan: one card per table, its columns and types, the key it would use and why, and how the tables join. Rename a table, retype a column, untick what you don't want. Nothing exists in your workspace until you press Import.

Once loaded, every table is a card on the Map: its rows, its key, its columns, and a dot that says how current it is. Relationships are lines between the cards. Before one is applied it is validated against the data — types, uniqueness, orphans — and when rows don't match you get three honest choices: fix the data, join without a constraint, or say it isn't a relationship. An invalid constraint is never created.

Every table is read for what it is: its domain, its grain (“one row per order line”), its measures, its time column, and anything sensitive — salaries, personal details, account numbers. Rules do this first, with no model at all; the model, when your workspace allows it, may add a sensitivity but never remove one. You confirm, edit or ignore the result. It labels things. It never grants anything.

Give a role a name and a sentence. You get the tables it should probably see, grouped by confidence, each with its reason. Sensitive tables are never pre-ticked unless the sentence names what they hold. Save grants exactly what is ticked; until then, a new table is the admin's alone.

Pick a role and the whole workspace answers as that role would — the shelf, the dashboards, the questions and the refusals — checked in the database, not merely hidden on screen. An amber line says whose eyes you're borrowing. It ends when you say Stop, or after an hour.