Shipping document verificationSI vs draft BL · APRIL operations inbox
520 emails in 3.9s · no LLM calls AI no key

Upload new data

Nothing here is written into the bundled dataset, and you can put it back with one click.

Compare one pair

.docx, .pdf, .txt, .xlsx · up to 8 MB each.
The AI fallback is currently off, so anything the rules cannot parse will be escalated rather than read by Claude.

Load a whole inbox

inbox/email_*.json — each with email_id, from, subject, body and attachments
attachments/ — the files those records point at, named <id>_SI.* and <id>_BL.*

currently 520 emails loaded

What happens to an uploaded document

The same code path the bundled inbox takes — no special case. The rules read every label they recognise, aligning by meaning rather than by header text, so Load Port and Port of Loading land in the same field. Anything they cannot parse, or a scan with no text layer, goes to Claude; anything neither can settle is escalated with the evidence rather than guessed.

An answer is cached against the exact bytes of the request, so running the same documents again is free — but the first pass over anything new is a real call.