ML Pub Club #21: Designing Models and Scalable Systems for Music Source Separation

​We live in the era of large language models — but what happens when you're working on a problem where you can’t just fine-tune a massive pretrained model?

​At our ML Pub Club, Stipe Kabić, Machine Learning Engineer at Atomic Intelligence & Daniel Vusić, Software Engineer at Atomic Intelligence,  will take us into the domain of music source separation - the task of isolating vocals, drums, bass, and other elements from a full audio track. From understanding state-of-the-art research to developing custom model architectures and building scalable systems around them, this talk covers the full journey of creating a real-world ML-powered product.

​In this talk, you’ll learn what music source separation is and how it works, how cutting-edge research in other domains can inspire better model design, how to go from training to efficient, scalable inference, how to architect a robust system for dynamic ML applications, and a look under the hood of StemNJam, an AI-powered music platform.

​Whether you're a machine learning engineer, audio tech nerd, or just want to learn how cutting-edge research turns into production-ready tools, we’ll see you May 6th!