Airflow in Action: Orchestrating AI At Duolingo To Unlock 99% Lower Costs, 10x User Growth

Belle Romea, a Software Engineer at Duolingo, discussed the creation of DuoFactory at the Airflow Summit, showcasing how the company leverages Apache Airflow to orchestrate large-scale generative AI workloads. Duolingo's need for a unified orchestration platform arose from fragmented tools and independent approaches to integrating LLMs into production pipelines, resulting in increased tech debt despite reduced content generation costs. DuoFactory, built around Apache Airflow, addressed these challenges by providing scheduling, dependency management, and integration with existing infrastructure, allowing modular and reusable pipelines that enhance efficiency and scalability. This system significantly reduced production time and costs for features like Duoradio, a listening feature, by automating script and exercise generation, which increased output and user engagement. Tools like the Prompt Editor and Google Sheets inputs broadened accessibility to non-engineers, allowing them to manage workflows without coding knowledge. Duolingo's experience demonstrates that while LLMs facilitate content generation, orchestration through Airflow is crucial for scaling and automating complex AI workflows, ultimately improving learner reach and content consistency across courses.