Every year, Spotify Wrapped gives hundreds of millions of listeners a snapshot of your year in listening. For 2025 Wrapped, we wanted to go deeper. What if we could identify interesting listening moments from your year, and tell you a story about them? The day you discovered that artist that changed everything. The day you played nothing but “yearning” music for six hours straight. The day your music taste took a hard left turn.

The Spotify Engineering Blog provides  insights, technical deep dives, and behind-the-scenes stories from the engineering team at Spotify. Covering a wide range of topics, including backend systems, data processing pipelines, and frontend architecture, the blog offers a glimpse into the technology stack and engineering practices used to power one of the world's leading music streaming platforms. Developers can learn about scalable infrastructure solutions, innovative engineering techniques, and best practices for building high-performance software systems. By following the Spotify Engineering Blog, developers can gain inspiration, knowledge, and practical insights to apply to their own projects and endeavors.

Spotify Labs

Spotify's engineering team details how they built Wrapped Archive for 2025, a feature that identifies up to five remarkable listening days per user and generates personalized LLM-written narratives about them. The system processed around 350 million eligible users, pre-generating roughly 1.4 billion reports over four days. Key technical challenges included designing heuristics to surface meaningful days, prompt engineering for creative consistency, distilling a frontier model into a smaller production model using DPO and curated human preference data, handling concurrent writes safely via column-oriented schema design, pre-scaling infrastructure before a simultaneous global launch, and building an automated LLM-as-a-judge evaluation framework to catch failures across the billion-report corpus.

Inside the Archive: The Tech Behind Your 2025 Wrapped Highlights

How we taught an algorithm to spot your big day

Yes, we actually generated 1.4 billion reports

Lessons in scale, safety, and storytelling

<p>Very nice ideas here for next year’s Log 😉</p>
