Technology

There is no recommendation engine that is right for every catalogue. A packaged one gets a shop running quickly, and for plenty of retailers that is the correct first move. What it cannot do is account for the things that actually decide which algorithm wins: how much sales history sits behind a typical product, whether shoppers arrive identified or anonymous, how fast the offer turns over, how hard seasonality bites, what your competitors are doing to the same customer on the same day. Those vary between retailers. They also vary between product groups inside a single retailer, which is why a shop where footwear behaves nothing like beauty is ordinary rather than unusual.

So MLArt does not ship one model and ask your catalogue to accommodate it. We read the data first: catalogue structure, what the click and view history actually covers, how the offer moves week to week, where the seasonal edges fall. That reading sets the configuration. It decides how much work reinforcement learning does, where language models over product text earn their place, where image models carry more signal than words, and where a well-tuned collaborative filter is simply the right answer and nothing exotic is needed.

The engine underneath is five modules. They are built to run together, and how they are weighted is the part we tune for you.

Five modules, one engine

Reinforcement learning. Contextual bandit algorithms choose what to show while they are still learning what works. Every impression does two jobs at once: it is a recommendation, and it is an experiment that improves the next one. That balance between using what the model already knows and testing what it does not is what lets the engine cope with an offer that changes daily, a product launched this morning with no history behind it, and a page whose content is different from yesterday’s. The contextual half conditions each choice on what is known at that moment about the product, the shopper and the page.

Natural language processing. Product descriptions carry information that no click log contains. The NLP module finds similarity in that text, which matters most exactly where behavioural data is thinnest: a new item, a short sales history, stock migrated in from another platform. It also lets recommendations follow editorial content, so a product block sitting under an article reflects what the article is about rather than what sold best last week.

Visual recognition. In categories where the look is the product, images decide similarity better than any field in the feed. This module matches on the image itself and generates tags from it, which turns a thinly described catalogue into a richly attributed one without a manual tagging project. The attributes it produces feed everything downstream, including search and filtering.

Classic recommendation. Collaborative filtering, content-based recommendation and basket analysis. These are long established and, for a stable catalogue with dense purchase history, still hard to beat. We treat them as both a working component of the engine and the baseline any newer approach has to outperform before it ships.

Connector. The module that talks to your systems. Catalogue, product feed, orders and live events come in; recommendations go back out to the storefront, the app or wherever else you serve them. It is unglamorous and it is the piece that decides whether the rest of it ever reaches a shopper.

A sixth module, added since

The five above were the engine in 2022. Since then a sixth has joined them: a generative and agentic module that gives the stack a way to speak and act.

It retrieves from your live catalogue, stock and policies before it generates anything, so answers are grounded in your records rather than in whatever a language model recalls about your category. It writes merchandising copy from the attributes already attached to a product, including the ones the visual module produced. And it runs bounded agents for post-purchase tasks, with irreversible actions stopping for a person. The full picture is on the generative commerce page.

Deployed in your environment

All of it is built and deployed inside your own infrastructure. Your product data, your customer events and the trained models stay where your other systems live, under your access controls and your retention rules.

That is a deliberate difference from renting a black box. You can see what the engine is doing, change it, and keep running it if our engagement ends. The configuration is yours because the deployment is yours.

See what it delivers on the solutions page, how it performed in practice in the case study, or email office@mlart.biz.