Customer Success Story
2
customized projects, four to six months apart
30K
products with outfit recommendations
90K
outfit combinations regenerated as the catalogue turns
0 hrs
manual styling where every look was previously built by hand
Founded 2017, Warsaw. Owns Born2be, Renee and Vices.
115,000 SKUs in active sale across women's, men's and kids'.
10 million orders and 250 million page views in 2024; 20,000 shipments a day.
Poland primary, Romania for Born2be and Renee.
Builds its own software — e-commerce platform, warehouse management, product catalogue, logistics, analytics team, in-house photo studios.
AZA Group came to YesPlz with a tagging problem and came back four to six months later with a harder one. The first project rebuilt their attributes around their own catalogue and taxonomy. The second used them — partly — to build three complete looks for every one of 30,000 products, each one passing a styling check before it ever reached a shopper.
Project 01 — Tagging built to their catalogue
The problem. Generic fashion taxonomies describe a generic catalogue. AZA merchandises on their own categories and on dimensions a standard ontology doesn't model — silhouette from neckline to skirts type, situation, season, occasion. Attribute quality decides everything downstream: search, filtering, and anything built on top.
What we built. Tagging tuned entirely to AZA's catalogue and taxonomy rather than mapping AZA onto ours.
Why it mattered beyond search. The attributes aren't a search dependency — they're a data layer. The same situation and season tags are available to merchandising and campaign teams, which is what makes catalogue-wide personalization possible rather than page-level relevance.
Project 02 — Outfit recommendations
The problem, stated honestly. An outfit is not a tag lookup. Two garments can share every attribute and still look wrong together, because compatibility is aesthetic rather than logical. And the search space is brutal: 30,000 items is billions of candidate pairings before anyone asks whether the result is any good. This is why a tagging vendor can't bolt on outfits and a recommendations vendor without fashion-native attributes can't either.
How it actually works. Four stages
Represent — visual embeddings capture how a garment actually looks, alongside the attributes from project 01. Tags narrow the field; they don't decide what goes with what.
Generate — outfits are built across thousands of situations, not a handful of templates.
Evaluate — every candidate look passes a styling test for silhouette, colour and occasion before it can ship. This is the step a stylist does last and a rule engine never does at all.
Personalize — the system learns from shopper behaviour, tuning looks toward the personas and lifestyles actually present in AZA's traffic.
What shipped. Three complete looks for every one of 30,000 products, spanning day, night and weekend.
Performance measurement is underway; results will be published here.
