On April 17th, Superlinked CEO and co-founder Daniel Svonava joined the AI User Conference Developer Day with not one but two sessions packed with practical advice for building real-world vector search systems.
The day started with a hands-on workshop for developers and ML teams who wanted to make their search actually convert. Daniel opened by tackling a common frustration. You’ve got a vector database, maybe some natural language search wired in, but the results just don’t deliver. He walked through how to fix that by designing search for complex items like travel listings, job boards or fashion products with lots of metadata. He showed how to support natural language queries like “vintage sweaters under $200” or “highly rated hotels in Manhattan with good wifi” that match real user intent.
Then he moved on to real-time personalization. Daniel showed how to set up a feedback loop using click data, so search results improve every time a user interacts. The outcome is better ranking, more relevance and stronger product discovery.
He wrapped up the session with a case study from a fashion retailer that added $15 million in incremental revenue using Superlinked’s custom embeddings. They combined all their first-party data signals into a single search and discovery system that actually surfaced what users wanted, not just what vaguely matched the query.
‍
‍
‍
‍
Later in the day, Daniel gave a second talk titled “Stop using pre-trained vector embedding models.”
He started by looking at how large tech companies really handle vector search behind the curtain. Instead of relying on generic models, they train custom embedding models from scratch for every single use case. These aren't just fine-tuned models. These are fully tailored encoders built specifically for one job. Daniel explained why they do this and what it means for everyone else.
If you don’t have a research lab of your own, should you be doing the same thing? Or is there a smarter way to get most of the benefit without the cost? The talk focused on how to make that decision with real evaluations and metrics. Daniel laid out practical ways to test what works, compare trade-offs and pick the right approach based on your actual product goals.
Both sessions hit home. One room was full of teams trying to fix product search, the other had engineers knee-deep in vector infra. A lot of people stuck around afterward to share what they were working on and ask questions about how to take things further.
Big thanks to the AIUC team for organizing a tight and useful event. If you're building anything in this space and want to chat, we’re always happy to connect!
‍