Building ML systems in business settings isn't just about algorithms; it's also about navigating unpredictable data, fostering efficient collaboration, and promoting constant experimentation. This talk delves into Metaflow's solutions to these challenges, spotlighting its event-driven patterns that streamline model training while leveraging infrastructure shared by multiple data-scientists/engineers. Through a hands-on demonstration, we'll showcase how Metaflow facilitates asynchronous training upon new data arrivals for finetuning a state-of-the-art LLM like LLaMAv2 (although the framework is general and widely applicable). Additionally, we'll highlight how multiple developers can seamlessly experiment with various models as new data becomes available.