Leonora Vesterbacka

Founding Data Scientist at Redpine, Stockholm.

Leonora Vesterbacka is Founding Data Scientist at Redpine, Stockholm. She holds a PhD in Physics from ETH Zürich, spent six years in research at CERN, and worked on applied machine learning as a postdoctoral researcher at NYU. She led the development of KB-Whisper, the best Swedish speech recognition model at the National Library of Sweden, and now works on unlocking non-public and paywalled data for AI at Redpine, building infrastructure that fairly compensates data owners rather than relying on scraping. She leads the scientific endeavour at Redpine, from developing rigorous evaluation methodologies to exploring training data attribution techniques.

Abstract:

At CERN, I worked with petabytes of data every day. Six years of particle physics research taught me one blunt lesson: the science is only as good as the data behind it. That lesson followed me out of academia. When I moved to the National Library of Sweden to build KB-Whisper, a Swedish speech recognition model, what made it good wasn't a clever architecture, it was fifty thousand hours of well-transcribed Swedish speech that simply hadn't existed as a usable dataset before. At Redpine, I now work on the other side of that same problem: most of the world's data is outside of the internet entirely, so we build the infrastructure to unlock it for AI, and make sure the people who created it actually get paid. This talk follows that thread, from CERN to KB-Whisper to Redpine, and why data, not models, is usually the real bottleneck.