Service 02 · Catalog
AI catalog enrichment.
Turn product data into structured data. I build pipelines that extract attributes, normalise taxonomy, write descriptions, and localise catalogues — without losing SKU truth.
- Best for
- Multi-store catalogues
- Pilot
- 4–6 weeks
- Approach
- Sample before scale
- Review
- QA sampling + queue
The problem
Product data arrives from suppliers in inconsistent formats. Attributes are missing, specs live in PDFs, and the same part has three names across two storefronts. A model asked to “fill in the gaps” invents specs and puts them in front of customers.
The work is not one prompt. It is a pipeline with source-of-truth rules, extraction limits, sampling, and a review queue that catches the long tail before publication.
What I build
Attribute extraction
Structured fields from titles, PDFs, images, and supplier feeds.
Taxonomy + dedupe
One category tree, canonical product names, and merge rules.
Descriptions
Copy generated from verified attributes, not invented claims.
US / Canada localisation
Spelling, units, pricing context, and market-specific copy.
Review queue
Low-confidence items routed to a human with the source attached.
Bulk pipeline
Batch processing, retries, cost per SKU, and export to your PIM.
How it runs
-
01
Catalog audit
1–2 weeks
Field coverage, duplicate rate, source quality, and the first schema.
-
02
Sample batch
1 week
Run 200–500 SKUs, measure accuracy field by field, and tune rules.
-
03
Pipeline
2–3 weeks
Batch processing, confidence scoring, review queue, and export.
-
04
Handover
1 week
Schema, prompts, QA dashboard, and a runbook for new suppliers.
What you get
- A documented attribute schema and taxonomy rules.
- Extraction and copy prompts with source citations and confidence scores.
- A sample QA report with per-field accuracy.
- A review queue for low-confidence SKUs.
- Export to your PIM, storefront, or CSV, with cost per SKU tracked.
- A runbook for onboarding new suppliers and markets.
Guardrails: no invented specifications, source attached to every generated attribute, human sampling before publication, and a hard cost ceiling per batch.
Questions
How accurate is attribute extraction?
It depends on source quality. The sample batch measures per-field accuracy before you commit to a full run.
Can it handle thousands of SKUs?
Yes. Batches run with retries and cost tracking; low-confidence items route to review instead of publishing.
Which languages and markets?
English first, with US and Canada variants. Other markets are scoped separately.
Start
Book a fixed-scope audit. You leave with a schema proposal, a sample plan, and an accuracy baseline.