Guards which AI actually works at SyncRefine, and the process that keeps enrichment reliable at scale.
LinkedInJulian Marulanda focuses on AI innovation and process management at SyncRefine. He tests which new AI models and techniques genuinely perform better on product data, and guards how they land safely in the platform: a sample first, only then the whole catalogue, and always with a person approving the proposals. The other half of his work is the process around it. He standardises the onboarding of new suppliers and audits the enrichment and publishing flow, so anomalies are caught before they reach a customer. More than once an improvement in the platform came straight out of such an audit: a duplicate row, a mis-mapped field, a piece of copy that was subtly wrong. On this blog he writes about AI in practice: what new models can and cannot do, how to use AI reliably at scale, and how a tight process keeps large assortments consistent and low on errors. His take: AI speeds up the work, a good process makes it reliable.

Nordic B2B is betting on AI, but AI scales bad product data instead of fixing it. What CDON, PriceSpy and PriceRunner reject, and the layer that fixes it.

Your supplier lists every size and color as a separate product. How to merge simple products into configurable or variable products, without broken URLs.

Product data quality starts with a gate, not a clean-up: quality gates block pricing errors, duplicate SKUs and mass deletions before they go live.

Product image optimisation done right: why heavy supplier photos wreck your LCP and Core Web Vitals, and how WebP and the correct size make pages fast.

Winning on CDON is not a translation job but a data one: EAN, MPN, brand and a real localized description per Nordic market, Finnish and Nynorsk included.

Product data consolidation done right: deduplicate supplier feeds on SKU and EAN into one clean golden record, with per-field provenance you can trust.

AI product descriptions only work when the product data underneath is clean. Write unique, findable copy, fill in attributes and translate into 42 languages.
Book a no-obligation demo and we will show it live on your own data.