Spotify Engineering builds and scales the technology behind the world's most popular audio streaming platform, serving over 500 million users. The team is known for pioneering open-source tools like Backstage, an internal developer portal used by 2,700+ engineers to manage 14,000+ software components, and Voyager, a high-performance nearest-neighbor search library for audio recommendations. It also created Klio for scalable audio data pipelines and XCRemoteCache to slash iOS build times by 70%. These innovations demonstrate Spotify Engineering's commitment to developer experience and data-driven audio innovation.
Spotify has introduced Random Access Parquet (RAP), an indexing method designed to enable low-latency point queries directly on existing data lake files. By mapping keys to specific file locations, the approach allows streaming services and AI agents to retrieve user data without the overhead of traditional distributed SQL engines or separate key-value stores.
Spotify has introduced Vedder, an agentic AI data assistant designed to provide reliable SQL queries across its 70,000 datasets. The system utilizes human-curated context clusters and domain expert validation to ensure data accuracy, rejecting 87.5% of automated query history as noise.
Spotify has released an open-source Claude Code plugin that enables advertisers to manage campaigns on the Spotify Ads API using natural language. The tool utilizes Markdown-based agents and the OpenAPI v3 specification to automate multi-step campaign orchestration without requiring traditional SDKs or the Model Context Protocol.
Spotify reports that AI-assisted development has doubled its monthly code changes to 17,000 while maintaining stable rework rates. The company is implementing enhanced monitoring, capacity planning, and automated safeguards to manage the increased velocity and prevent regressions in its content processing and mobile application infrastructure.
Spotify researchers evaluated the efficacy of using LLMs to replace human participants in A/B testing, finding that raw LLM predictions captured only 39% of human treatment effects. The study concludes that while LLMs can serve as useful proxies for variance reduction or filtering, they cannot replace human experiments for novel interventions due to untestable assumptions regarding surrogacy and comparability.
Spotify Engineering details the architecture of "Vedder," their internal AI data assistant, which helps employees query petabytes of data across 70,000 datasets by generating SQL queries. The system relies on human-curated data domain "clusters" and explicit expert review of training examples to ensure accuracy and trustworthiness, with experts accepting only 12.5% of automatically proposed question-SQL pairs. Vedder has been used by over 2,100 Spotify employees in over 13,000 conversations since August 2025.
Spotify reports that 99% of its engineers use AI coding tools weekly, leading to a 76% increase in AI-authored pull requests. The company developed 'Honk,' a background coding agent powered by Claude, to handle complex code modifications within its Fleet Management system, significantly accelerating migrations. This shift highlights a change in development bottlenecks from coding to human decision-making, with tools like Backstage providing standardization for both human and AI agents.
