How Albatross deployed real-time perception models across Wallapop — a second-hand marketplace with more than €2 billion in GMV — and rebuilt discovery on five surfaces in two quarters.
× 
A summary of the Wallapop × Albatross collaboration through Q1 and Q2 2026.
Wallapop is one of Europe's largest second-hand marketplaces — more than €2 billion in annual gross merchandise value, and a catalog that turns over by hundreds of thousands of listings every day. That mix of scale and churn is exactly what conventional recommendation systems handle worst.
Over two quarters, Albatross replaced Wallapop's legacy discovery stack with a set of real-time perception models. A small, shared set of primitives now powers five distinct surfaces: the personalized home feed, a purchase-optimized variant of it, the similar-items carousel on every listing, a new playlist-completion product, and personalized marketing outreach. Each was built and A/B tested.
At Wallapop, we are moving toward a system that understands what users want in real time, helping buyers find the right items faster while giving sellers more effective visibility for their listings. Our collaboration with Albatross represents another step forward in our mission to empower people to participate in a more conscious consumption model that creates economic opportunities for people.
We typically don't outsource systems this central to our business. That we're working with Albatross says a lot. Their discovery engine allowed us to launch a completely new experience in weeks, not years, while significantly improving the user experience.
Fast-changing catalogs of unique items defeat collaborative filtering and rules.
For any marketplace, the core recommendation challenge is structural. The catalog changes extremely rapidly and consists of unique items with inconsistent metadata. Hundreds of thousands of new listings enter the catalog every day with no impressions and no interaction history — which means traditional systems cannot recommend them at all.
Collaborative filtering and rules-based engines were built for stable catalogs of repeatable products. They struggle on two fronts at once. They cannot keep up with fast-changing user intent: recommendations stay static as a person browses, clicks, and refines what they are looking for. And they cannot keep up with a dynamic catalog — a legacy system typically needs a sustained volume of interactions before it will recommend a new item with any confidence, so fresh inventory can take days to surface.
A marketplace only works if sellers keep listing items — and sellers only keep listing when those items surface and sell.
At this size, small gains and small losses both become very large numbers.
Scale is what turns this problem from an inconvenience into a material cost. Wallapop operates one of Europe's largest second-hand marketplaces. Annual gross merchandise value exceeds €2 billion. The live catalog runs to roughly 60 million items. Hundreds of thousands of new listings are added every day, and the marketplace spans multiple countries, including Spain, Italy, and Portugal.
Sequential embedding models trained on live user-event streams.
Albatross takes a different approach. Instead of collaborative filtering or hand-tuned rules, it uses sequential embedding models trained on live user-event streams. These models learn how users respond to products, content, and discovery decisions, then predict the next best items or actions to serve at each point in the journey.
The defining property is that the model updates in real time. Every new click updates Albatross's understanding of the user immediately. Recommendations are regenerated on the spot and served with low latency, so the feed adapts as the user explores rather than staying fixed for the session. A new listing can become recommendable within hours of entering the catalog instead of waiting days for clicks to accumulate.
Between June 2025 and May 2026, the collaboration moved from a single pilot to five discovery surfaces, four of them live in production.
The home feed is the entry point of Wallapop. With Albatross, every user sees a unique, personalized list of items that refreshes after every click — a real-time discovery experience that evolves as the session unfolds. The pilot was evaluated over a long run; the signal was unambiguous, and the feed has since expanded internationally.
More than a doubling of engagement against the system it replaced.
The home feed is tuned for engagement and discovery. The purchase-optimized feed is the same feed reranked toward a different objective: completing a sale. It pairs the Discovery Feed primitive with Dynamic Reranking, which reorders the feed's candidate set against an explicit purchase objective. Launched in April 2026, Albatross's funnel analysis showed the reranked feed moved users further down the purchase funnel without sacrificing discovery.
On every item-detail page, a carousel shows the best alternatives to the listing a user is viewing. Before Albatross, Wallapop's in-house solution covered less than half the catalog — most listings simply had no good alternatives to show. Albatross delivered the endpoint, ran an A/B test, and rolled it out globally. Coverage rose from under 50% to 94%, and the carousel became one of the highest-traffic surfaces on the platform.
Smart Playlist Completion is a new product — the first of its kind on Wallapop — built on the same feed technology. Users create lists of favorite items, and Albatross recommends the ideal items to complete each list. The recommendations are personalized twice over: unique to the user and unique to the list, so the same person receives different suggestions for different playlists.
Recommendations regenerate with real-time context — unique per user and per list.
The fifth deployment extends personalized discovery beyond the app itself. Marketing Outreach reuses the Discovery Feed primitive to bring the same real-time, personalized item selection into outbound marketing channels.
One primitive, a second surface. The Discovery Feed that powers the home feed needs no rebuild to drive outbound channels — the same building block is simply called from a new place.
Across the past six months, all surfaces have sustained an SLA of 99.8–100%, with 95th-percentile latency under 90 milliseconds.