VeloDev
/ AI product · WooCommerce

Catalogue clean-up, with your approval on every change

Most store catalogues carry years of copy-pasted supplier text, English descriptions and empty fields. AI Product Manager finds them, proposes better names and descriptions written only from the product's real data, and changes nothing until the owner approves. Every published batch can be reverted.

Platform
WooCommerce
Status
Working MVP · open for pilots
Writes to the store
Only after approval
AI cost per product
~$0.0003
Technology Stack
  • Node.js
  • Fastify
  • TypeScript
  • PostgreSQL
  • React 19
  • Vite
  • Claude API
  • WooCommerce REST API
  • CSV import/export
/ demo

The full flow on a demo store: import the catalogue, see the issues, generate proposals for four products, approve them one by one and publish. The proposals are the model's own output, unedited.

product-manager / dashboard
Demo
Demo store in English · Claude Haiku · 4 products, 12 proposals, $0.0015 · 2.5× speed · the batch was reverted afterwards
Launch test on the demo store
60
Changes published
60/60
Restored exactly by revert
20/20
Products passed the AI eval, EN and RO
17
Issue types detected
20 products × 3 fields with Claude Haiku: $0.0068 in total.
/ what it does

A list of what is wrong

17 checks run on the whole catalogue for free, with filters and search, so the owner sees where the work is before spending anything. Works for English and Romanian stores, including text in the wrong language.

Text from facts only

Proposals use only the product's own data. Unverified numbers or claims get a warning, and missing facts are listed instead of invented.

Nothing changes without approval

Approve, reject or edit each proposal. Publish directly to the store in small batches, or download a CSV for WooCommerce's own importer.

One-click revert

Every batch keeps the old text. Products edited in the store meanwhile are skipped, never overwritten.

/ Engineering snapshot

Challenge

Rewrite product copy at catalogue scale without the two risks that make owners say no: invented product claims, and bulk changes that cannot be undone.

What I owned
  • Import from WooCommerce REST and CSV, with identical results from both.
  • The rule-based detector and the AI proposal jobs with cost caps.
  • Review, publishing, history and revert, including conflict handling.
  • Prompt and evaluation set, tuned against a real model.
Result
  • 60 changes published to the demo store in batches of 10, then reverted with 0 differences.
  • 20/20 products pass the eval in English and in Romanian; the hidden-instruction product never produced an invented claim.
  • About $0.35 of AI cost per 1,000 products with Claude Haiku.
Key engineering decisions
  1. Detection without AI

    Finding problems is deterministic and free; the model is used only to write proposals for the products the owner selects.

  2. Fact guard and injection filter

    Sentences that try to instruct the model are stripped from the source, and any number or claim not found in the product data is flagged before review.

  3. Optimistic conflict check

    A change is written only if the store still holds the text seen at import; revert does the same check in reverse.

  4. Tuned on what a reader notices

    The first real-model run passed the fact checks but read badly (filler phrases, SKUs in names, missing diacritics). The prompt was tightened and the eval now checks each of those.

/ screens

Product Manager screens (4)

Click a tab to switch, or the image to enlarge.

product-manager / dashboard
01 / 04

Catalogue issues

After importing from the store (or a WooCommerce CSV export), a rule-based detector lists 17 kinds of problems: missing descriptions, ALL-CAPS names, English text, supplier codes, messy HTML, inconsistent attributes and more. No AI, no cost.

/ how it connects

Import

  • WooCommerce REST API or the standard CSV export.
  • Imports kept in PostgreSQL with full history.
  • Rule-based detector, 17 issue types.

AI

  • Background jobs of up to 50 products.
  • Cost cap per job and per month.
  • Fact guard and prompt-injection filter.

Publish

  • REST batches of 10, or CSV plus a rollback CSV.
  • Conflict check before every write.
  • Revert per batch or per change.
/ pilot

Try AI Product Manager on your store

The product works end to end on a demo WooCommerce store; the next step is a pilot on a real one. A pilot is a few weeks on your store with your data, read-only or behind your approval, with the results reviewed together before anything is extended. Agencies can offer it white-label to their own WooCommerce clients.

marian.putirac@gmail.com