Project

Clean-Data

Paste a messy order and get back line items matched to the catalog. Nothing uncertain goes through until a person signs off.

Open the live demo Source
Clean-Data’s order review for a forwarded text message: 20 line items, 19 matched to the catalog and 1 flagged for a person, with 9 typos fixed and 28 abbreviations decoded
19 of 20lines matched in a forwarded text message
28abbreviations decoded in that message, plus 9 typos fixed
0accounts or keys needed to try it

In motion

The edge case, end to end: one forwarded text message in, 20 catalog lines out.
Clean-Data, live demoOpen in a new tab: Clean-Data, live demo
Open the live app

Once it’s open, press Run Demo, then try the Edge Case.

Why I built it

At Ventura, we sold industrial distributors an agent that reads their orders, down to an email that says “blk iron niples 3/4”. The product was hard to explain in a sentence and clear after ten seconds of watching it work, so I built a version anyone can run. It still carries Ventura’s name in its header.

What I decided

  • The send button waits for a person. Anything it can’t place is flagged, and sending stays locked until someone has accepted or rejected each one. No setting turns that off.
  • Show the receipts. Hover any row to see the exact words it was read from. That’s how a distributor’s rep learns to trust it.
  • Zero setup. A deterministic parser runs the whole demo offline.
A spreadsheet export full of typos, read into 20 catalog lines with a confidence meter on each, every one of them matched
Another sample, a spreadsheet export full of typos: all 20 lines matched.

How it works

  1. An order, however it comes inEmail, purchase order, spreadsheet or text message; with a key, a photo too
  2. One parse endpointThe same route for every input

  3. Is an API key set?

      1. A hosted modelWith the catalog in its prompt
      1. A deterministic parserSeven extraction patterns, offline
      2. A fuzzy catalog matcherBy aliases and word overlap
  4. Line items, matched to the catalogEach with a confidence score
  5. A person reviews the orderAccepting or rejecting each flagged line
  6. Sent to the ordering system (ERP)Or exported as a spreadsheet (CSV)
Either engine can read the order. Only a person can send it.