About MLArt
MLArt builds AI systems for online retail: recommendation, ranking, search, pricing, and the generative layer that now sits over them. What we build is deployed inside the client’s own infrastructure rather than rented back as a hosted service.
The company has had two eras. The second did not replace the first, and that is the part worth explaining.
Built to survive a moving catalogue
MLArt began as a team of data scientists with a specific complaint about standard recommendation engines. They learn from history, then assume tomorrow will look like yesterday. In a shop where the promoted offer changes daily and new products arrive with no sales record behind them, that assumption fails immediately, and it fails hardest on exactly the products a retailer most wants to move.
What we built used contextual bandit algorithms, which keep learning while they serve rather than waiting for the next retraining run. Around them we put language models reading product descriptions and image models reading product photography, so an item with thin or no click history could still be placed on the basis of what it is rather than what it has already sold.
That engine reached production, delivered with a technology partner, replacing the recommender already running on a fashion e-commerce platform with a catalogue of more than a million products. The case study sets out what changed, including how products with no sales history were brought into recommendations at all.
Since: systems that talk and act
The newer half of MLArt is generative and agentic commerce. A shopping assistant that answers from your live catalogue, stock and policies. Merchandising copy generated from real product attributes and reviewed before it publishes. Agents that handle bounded post-purchase tasks, with anything irreversible held for a person to confirm. That work is set out on the generative commerce page.
None of it displaced the earlier work. A shopping assistant with no decision engine behind it is a conversation with no opinion: fluent, and unable to say which twelve products deserve the slot. So we kept the decisioning discipline of the first era and applied it to the second. Grounding is measured rather than assumed, generated output is checked against the record it came from, and an agent’s authority is written down before it is granted.
The company
MLArt Ltd operates internationally. We do not publish a registered address, a company registration number or a jurisdiction on this site, and that is a deliberate choice rather than an omission. Formal registration detail is available on request, and we provide it to anyone who needs it for due diligence or contracting.
Contact
One address reaches us: office@mlart.biz. Bring a catalogue, a conversion problem, or an existing recommender you suspect is underperforming, and we will tell you plainly whether we can improve on it.