Write a playbook of categories, decisions and rules, then drop in Excel, CSV, JSON or logs. It fills in each row, queues the unsure ones for you, and hands back the same file with new columns. Everything runs in the tab.
What it does
Rules always win. Then a word-based model, then embedding similarity, then an optional second opinion from a larger model on unsure rows.
A review queue with a reason for every row, and your corrections become training examples.
Excel comes back as the same workbook, with other sheets, formatting and formulas kept.
How it is built
- Measured on held-out lines: categories 93% of 29, ticket decisions 87% of 45.
- Mixing weights were tuned on separate lines before the test set was run.
- Models live in the browser's private file system and projects in IndexedDB.
Honest limits
- The test sets are tiny and hand-labelled.
- The larger model is 847 MB and needs WebGPU.