Solutions

MLArt’s core product is a recommendation and personalization engine for online retail. It reads your catalogue and your traffic, decides what to put in front of each shopper, and keeps deciding as both change. Three products make up the offer: predictive personalization, adaptive recommendations, and automatic product tagging. They share one data pipeline and one connector into your storefront, and they run inside your own infrastructure.

It is built for retailers and platform teams past the point where a packaged engine helps: catalogues too large to merchandise by hand, ranges that turn over faster than a quarterly retrain, or attribute data too thin for search and filtering to do their job.

Predictive Personalization

Personalization is a ranking problem with a deadline. A shopper lands, and within one page view you have to choose what to show from a catalogue of hundreds of thousands of items.

MLArt scores candidates against the context you actually have. For a returning, identified customer that includes view, search, click and purchase history. For an anonymous visitor, and most first sessions are anonymous, it uses what the session itself reveals: entry point, search terms, the category being browsed, the item on screen. Both paths run through the same model. Identity improves the answer; it is not a precondition for one.

The engine is useful on day one, working from catalogue structure and live behaviour, and it sharpens with every interaction after that. The same scores drive on-site placements, the mobile app and outbound messaging, so a customer meets one consistent view of what you think they want rather than three contradictory ones.

Adaptive Recommendations

Most recommender systems carry a hidden assumption: that next month will look like last month. Train on history, apply the model, refresh occasionally. In a stable catalogue that holds well enough. In practice, commerce moves constantly, and it moves in specific ways that break the assumption:

• the active offer changes daily as items sell out, return, or lose priority against the current sales target;
• new products arrive with no sales history at all, so nothing in the data says how they will perform (the cold-start problem);
• seasons shift demand, with some products selling only inside a window, some converting better in one, and some indifferent to the calendar;
• popularity moves, and occasionally a product becomes a hit faster than any retraining schedule registers it;
• customer preferences and behaviour drift.

Adaptive recommendations treat each impression as a decision made under uncertainty rather than a lookup against a frozen model. Every placement does two jobs: earn a click now, and buy information about products the system does not yet understand. That trade is managed deliberately, and it is what lets a product earn its place in a recommendation slot on its first day instead of its first month.

For the mechanics, including how contextual bandits handle the exploration-versus-exploitation trade and what happens to a brand-new product, see adaptive recommendations and the cold-start problem.

Automatic Product Tagging

Shoppers can only find what your catalogue describes, and most catalogues describe less than their owners think. Attributes arrive inconsistently from suppliers, taxonomies drift, and the fields that filtering and similarity depend on end up half-populated. Manual tagging closes that gap slowly, expensively, and differently depending on who does the work.

MLArt closes it with two readers working on the same product. An image model looks at product photography and recognizes what is visually there: category, colour, pattern, cut, material cues. A language model reads titles, descriptions and supplier copy for the attributes stated in words rather than shown in pictures. Together they produce one consistent attribute set across the whole catalogue, applied to new products in minutes.

Because it is a model rather than a team, changing your mind is cheap. Redefine the taxonomy, add attributes, re-run across everything.

Richer attributes pay off twice: facets and filters get better, and the recommendation models get the product signal they need when there is no sales history to lean on. More on the method in automatic product tagging.

Use cases

• Similar-product recommendation. Keep a shopper engaged when the item in front of them is close but not quite right.
• Complementary and cross-sell recommendation. Show what goes with what has already been chosen, on the product page and in the basket.
• Checkout-area recommendation. Add small, plausible items on the cart and checkout pages without getting in the way of the purchase.
• Content-aware recommendation. Match products to the article a reader is on, so editorial and blog traffic has somewhere to convert.
• Catalogue enrichment. Extend product attributes across the catalogue using image recognition and language models, and keep them consistent as stock turns over.

Which of these fits depends on your data and your catalogue dynamics, not on a product tier. Technology sets out the modules underneath: reinforcement learning, natural language processing, visual recognition, the classic collaborative and content-based methods, and the connector that moves data between your systems and the engine. The case study covers a 2022 engagement with a fashion e-commerce platform where this combination lifted click-through by around 30% and brought products with no sales history into recommendations. To find out which parts apply to your catalogue, write to office@mlart.biz.