New on the blog: how we infer user journeys - the latent goal behind weeks of saves, searches & closeups. Learn how we serve fine-tuned LLMs to hundreds of millions of users, and what's next with Semantic IDs & generative recommendations: medium.com/pinterest-enginee…

Learn about Metrics Board, our platform for defining, computing, and publishing metrics. It drives speed, context, and efficiency with one idea: define a metric once, as a versioned spec, and let everything downstream follow - including analytic agents. medium.com/pinterest-enginee…

We're excited to announce our 7th Annual ML Day Conference on November 5. Get ready for keynote sessions with Pinterest Engineering leaders, panel discussions on AI at Pinterest and lightning talks that dive deep on ML innovation. Register now: events.ringcentral.com/event…

Streaming ingestion gives us fresher data in minutes, but how do we know when a partition is complete enough to read? Learn how we restore that guarantee with Iceberg commits, safe-to-read watermarks and reusable Flink-to-Iceberg sink extensions. medium.com/pinterest-enginee…

Learn how the team launched the first 3-tower engagement co-train model for ads lightweight ranking at Pinterest, introducing richer user-ad interactions while maintaining the latency and cost efficiency required at scale. medium.com/pinterest-enginee…

Manas serves billions of embeddings across 80+ clusters at Pinterest, but scaling to tens of billions, we needed new approaches: quantization (50%+ memory savings), SSD-based serving (10x less RAM), & multi-embedding retrieval for richer ranking. Details: medium.com/pinterest-enginee…

Get to know the story behind building our new multimodal AI foundation with NVIDIA. Our team shares insights about enabling Pinterest Assistant to run 25x more visual context per request. medium.com/pinterest-enginee…

Today, @Pinterest announced a new foundation for multimodal AI built with @nvidia to support a growing range of products that rely on both images and language. Learn more: newsroom.pinterest.com/news/…

In our latest blog, we introduce Conditional Learned Retrieval, a powerful two-tower architecture combining state-of-the-art ML and modern RecSys methods to balance user engagement with contextual relevance to users' various interests. Learn how we did it: medium.com/pinterest-enginee…

🚀We break down RPP - our purpose-built terraform execution engine for securing tens of thousands of AWS resources across hundreds of workspaces. 🔒See how we combine OIDC role-chaining, backend checks, and PR dual controls to lock down multi-repo IaC 👇 medium.com/pinterest-enginee…

Calling all aspiring product managers, designers and researchers 📢 Applications for Pinterest’s 2026 Products Apprenticeship Program are now open. Apply now: pinterestcareers.com/departm…

Schema changes shouldn't break your offline DB ingestion pipeline. Our latest blog explores how Pinterest automates schema evolution for its next-generation DB ingestion framework, safely propagating schema changes across Kafka, Flink, Spark, and Iceberg. medium.com/pinterest-enginee…

Adding a second GPU node made training 5x slower. Our team shares how we moved foundation model training from a broken 0.2x baseline to 7.5x scaling by fixing networking, profiling, communication, and distributed topology. medium.com/pinterest-enginee…

Our next Pinterest Labs Talk in one week away! 🚨 UC Berkeley's Dawn Song will join us for a discussion on recent advancements in AI & LLM agents and the risks they present. Register now to attend: pages.beamery.com/Pinterest/…

Join us June 25 for our next Pinterest Labs Talk with UC Berkeley's Dawn Song for a discussion on AI & LLM agent advancements and key risks, moderated by Chuck Rosenberg. Register now 👉 pages.beamery.com/Pinterest/…

Real-time ad retrieval power-up 🚀 Hybrid inference combines sequential transformers + context to blend high-intent signals with long-term user history. Result: 3x–10x Recall@K, ~300% higher median candidate relevance, and measurable ROAS lift. medium.com/pinterest-enginee…

New on the Blog: For large-scale ML systems, moving less data can be just as important as computing faster. This post shares how we reduce unnecessary feature-transfer overhead to improve network efficiency in production. Read more: medium.com/pinterest-enginee…

From Determinantal Point Processes to sliding spectrum decomposition, here's how our team evolved multi-objective optimization at Pinterest to balance what people engage with and what keeps the experience inspiring. medium.com/pinterest-enginee…