Services // Search & Recsys

Search & Recsys

Search that finds the right product and feeds that keep people coming back. We build ranking, recommendations and personalization on your catalog and behavior data.

Embedding models · Learning-to-rank · Vector search · SmartFeed · Redis

Discuss a search project What does it cost

What search and recommendations do for a catalog

Search turns a query into the right product. Recommendations turn a visit into the next one. Both read your catalog and your behavior data, then rank what a person is most likely to want next. The work is ranking: given everything you could show, put the best few at the top.

A feed on a home or category screen rarely comes from one place. It blends personalized picks, promoted items, best-sellers and what is in stock nearby. We run a feed engine that fetches those sources at once, removes repeats, re-ranks the result, and serves consistent pages as a person scrolls. Blend ratios, pinned slots and A/B tests are configuration, so changing the shape of a feed does not mean a rebuild.

When this makes sense

You need this if

  • The catalog is big enough that people fail to find things
  • Empty results and exits after search are measurable
  • Behavior data accumulates: views, carts, purchases
  • A feed blends several sources and someone hand-codes the mixing today
  • You want to pin promoted slots or run A/B tests without a deploy

Not for you if

  • A hundred items: a good menu beats a search box
  • There is no analytics on what users do yet
  • One source fills the whole screen and never changes

Architecture

Several data sources on the left, a SmartFeed box that blends, dedupes, re-ranks and caches, and one scrollable feed on the right with promoted slots pinned in place

Many sources go in, one clean feed comes out. SmartFeed blends them by a rule, removes repeats, re-ranks and caches. Promoted items sit in pinned slots, and you change the mix or run an A/B test by editing config, with no rebuild.

Proof

METRO MarketsSearch relevance across a six-market wholesale marketplace: a hybrid engine that reads the meaning of each query, and a testing setup the team runs to keep improving it.Social network feedA feed backend with a composition engine on top: the business spins up new feeds and A/B tests in minutes, blending several sources into one page.Marketplace cold-startRecommends brand-new products from their description alone, so the long tail gets shown from its first day with no click history.

Use cases

A few ways search and recommendations show up in a product.

E-commerce Search A search engine tuned to your catalogue: indexing, query parsing, weighted EQL queries, synonyms, listings. Iterative tuning by your own metrics, with the vendor defaults left behind. Product Recommendations Recommendation systems and smart feeds that handle cold-start, scale to millions of users, and integrate with your product. Including the cold-start architecture we shipped for Factorymarket (millions of SKUs, no purchase history). Product

All use cases →

Knowledge map

How we think about ranking, feeds and personalization.

Conceptual

Every engagement starts with a baseline: what your current search or feed does today, measured on your own traffic. From there it is ranking, the sources that feed it, and the tests that tell you a change is real before it ships.

What does search or a feed cost?

Most search and recsys work lands in the Standard tier of our estimator. Ranges are cumulative, with the plus or minus 15% promise we use everywhere on this site. Marketplace scale can move it to Complex.

  • Working PoC about 8 weeks €20-28k typical
  • A system that ships about 16 weeks €41-55k typical
  • Production grade about 32 weeks €82-110k typical

A pilot that proves the ranking beats what you have today is the cheapest thing we sell. Marketplace-scale catalogs and strict latency cost more.

Open the estimator →

Have a project like this?

Discuss a project → Estimate it