11 slides · speaker notes · the deck · present it
Recombee — AI personalisation and search
Who we are in one sentence: an AI personalisation and search engine that runs at scale and efficiently — most of our inference runs on CPU today because for our tasks it gives the best performance per price, GPUs where they pay off — with a university research lab behind it. First what we sell and to whom: three products and ten domains. Then the research behind them, and then what we could do for you.
One engine: recommendations, search and insights
What we sell, from the bottom up. More than a thousand servers in four regions — the US, Canada, Europe and Australia. On them runs the Recombee cloud: one distributed engine with more than a hundred models that learn in real time, thirty thousand recommendations a second. And that one engine powers all three products. Recommendations: every row of a homepage, similar items, what to watch or read next, the emails and push messages. Search that understands what a query means and ranks the results by each user's taste. And insights: live KPIs for every placement, with A/B tests built in. How clients steer it and plug it in comes on the next two slides.
One engine for every placement — the client keeps the rules, we keep the models getting better.
A few steps to personalisation
How it works, the way our website draws it. A client sends two things: the catalogue — items and users, with their text and images — and the interactions, live as they happen and the history. Both flow into the engine in real time. For each placement the client sets up a scenario: what kind of recommendation, which logic, filters and boosters, constraints. Out come the homepage, search, newsletters, push, similar items, watch next — one engine behind all of them. And every result flows back into analytics, so the client sees what each placement earns.
Data in, in real time — a scenario per placement — personalisation out everywhere, and every result back into analytics.
Ten domains, one engine
Where we specialise. The top row is media and entertainment: FTV Prima's video, streaming platforms built by Diagnal — and DAZN in sport — The Telegraph in news, Audiomack in music, 9GAG for user content. The bottom row is commerce and marketplaces: retail, classifieds, real estate, jobs, travel and deals. Each number comes from a published case study on our website.
The same engine wins in media and in commerce — the domain changes the data, not the product.
Sport & live events: every fan's favourite platform
Sport is our spearhead in media. Homepage rows built for each fan — live matches, replays, the leagues and teams they follow. The model learns during the match, not overnight. Editors still steer: premium matches, sponsors, local leagues — and it stays one-to-one. Semantic search for queries like "goals this weekend". Personal highlight reels. DAZN runs this in more than two hundred markets, on every device.
Sport is where personalisation has to be live — and where we already run at global scale.
Video, news & music: a feed for every viewer, reader and listener
Media is our home, next to sport. Broadcasters like FTV Prima, video platforms built by Diagnal: a homepage built for each viewer, the next episode while the credits roll. The Telegraph, where a new article has to reach the right readers within seconds, and search and recommendations run on one engine — they measured us against a competitor in an A/B test: thirty-five percent more clicks. Audiomack, where recommendations tripled monthly plays. What they share: catalogues change by the minute, traffic peaks during live events, every device counts. That is why our models are sparse and efficient — they answer in milliseconds, and most of them run on CPU, where the price-performance is best for this work.
Media is where personalisation has to be fast — and efficient at every request.
Recombee Research: 12 papers accepted in 2026
A word about where the products come from. Recombee Research and RecombeeLab — our joint lab with the Czech Technical University in Prague — had twelve papers accepted this year: at RecSys, UMAP, CIKM, WWW and IJCAI, with two best-paper nominations at UMAP. These are pages of papers we have already published: an industry-track paper at RecSys on embeddings ninety percent slimmer, the editorial analytics we built with The Telegraph, linear autoencoders at industrial scale, work at WWW and ICML. We fund PhD students who work on our production problems, and we open-source what we can. Everything published is on recombee.com/research.
Peer-reviewed, public and ours — the published work is at recombee.com/research.
Semantic IDs: what the leaders publish this year — and our breakthrough
The hottest topic of recommender systems right now. Google started it with TIGER; YouTube now has Gemini write every video as a code. Spotify uses semantic IDs for podcast discovery, Meta in its ads ranking, Kuaishou serves a quarter of its traffic with one generative model built on them, Alibaba's maps predict your next place as a code. What they share: giant generative transformers that are heavy to train and to serve. We have a scientific breakthrough here that works markedly better than semantic IDs and those giant transformers — and it is efficient. The paper is under review at WWW, so that is all I can say today.
The leaders bet on semantic IDs — our breakthrough beats them, and efficiently.
Efficient by design: share meaning and experience, serve sparse
Why efficiency matters. We serve thousands of databases; a GPU at every request would make cost grow with traffic. So we share what is cheap to share: frozen encoders and one sparse dictionary, computed once per item on CPU, and the experience of thousands of A/B tests — a small meta-model that predicts which algorithm wins for which data. Serving stays sparse and linear — thirty thousand recommendations a second — and today most of our inference runs on CPU: not on principle, but because for these tasks CPUs give us the best performance per price. GPUs do what they do best: training and the heavy encoders. And a CPU is only cheap when it is well used — keeping it busy is our job, not the client's.
The best model for each database — on the hardware with the best price-performance.
What we could do for you
What this would look like for you. We would not start with everything: one placement, the one where you most want to see a difference. You send us the catalogue and the interactions; the models learn from your own data. Then an A/B test against what you run today, on the number you care about. If it wins, we grow placement by placement — all on one engine.
One placement first, one number that proves it — then the rest.
Thank you
Thank you. The case studies, the research and the documentation are all on recombee.com.