Team Superlinked made a strong showing at this year’s AI Summit in London. We brought the full package: a sharp live demo, a panel full of practical insights, and a raffle at our booth that may or may not have involved a Jaws-themed LEGO set.
Our co-founder Ben joined the panel Accuracy Architects: Strengthening the Backbone of Enterprise Data on the Data Excellence Stage. Alongside Samrah Kazmi, Nimrod Vax, Laura Hughes, and Maraim M., the panel dug into real-world challenges like privacy, compliance, data monetisation, and what it actually means to use data well.
Ben focused on a simple but often overlooked idea: using the right data beats using more data. It’s easy to assume that quantity will win, but in reality, flooding your system with irrelevant logs or noisy history can make performance worse.
He shared a few examples. One team tried training a support bot with a giant archive of old tickets, but it kept getting confused. Replacing that with a focused, well-tagged set of common questions gave far better results. Another recommendation system was drowning in millions of click and scroll events. Once the team narrowed it down to meaningful signals like purchases and cart adds, accuracy improved dramatically.
At Superlinked, we bake this discipline into our systems using an embedding-first architecture. Rather than dumping everything into a model, we carefully design how each piece of data gets turned into a vector. That means product descriptions, prices, and category tags are each handled in their native formats. It results in cleaner, more reliable outputs, because the model understands what each bit of data actually means.
‍
Over on the demo stage, Filip showed off our natural-language e-commerce system with a session titled Type & Shop: Natural-Language Search and Live Recommendations in E-commerce.
He demostrated queries like “orange striped women’s sweater under 100 dollars”, showing how the Superlinked framework delivers fast and accurate results when faced with plain sentences. The system understands intent and filters at the same time. Each interaction, from clicks to add-to-carts, updates your session profile in milliseconds. The result is a feed that evolves in real time, complete with personalised product suggestions, discounts, and related items.
It is a real example of what happens when you combine high-quality data with the right architecture. No patchwork of filters or stitched-together models, just a clean flow from user query to result.
‍
Back at the Superlinked booth, we had plenty of great chats with data teams, AI leads, and curious passersby. The raffle kept things lively, and our team shared ideas on how embedding-first search can solve messy problems across verticals.
This year’s AI Summit was a chance to connect, share, and learn. We showed what embedding-first architecture can do, and how smart data selection can make or break AI outcomes. If you didn’t catch us at the booth or on stage, you can still reach out or check out the open source for yourself.
‍