AI BriefingPinterest Engineering news
The latest AI and developer news from Pinterest Engineering, with short summaries.
Visit website →Network Efficiency for ML Workloads (Part 1): Feature Trimmer
Pinterest reduced network bottlenecks in root-leaf ML serving by trimming unnecessary features.
From Clicks to Conversions: Designing Pinterest's Shopping Conversion Candidate Generation
Pinterest redesigned its conversion-focused candidate generation model, boosting RoAS and conversion performance.
Pinterest's MIQPS Achieves Large-Scale Content Deduplication with Smarter URL Normalization
Pinterest improved content deduplication accuracy by refining URL normalization with MIQPS.
Finding a Zombie in the System: A Real-World Case of a CPU Bottleneck
The cause of Ray training failures wasn't the GPU, but CPU starvation affecting the ENA driver.
Performance for Everyone
By embedding Visually Complete into BaseSurface, 60+ surfaces are measured automatically.
Pinterest Home Feed's Multi-Objective Optimization Evolution
Pinterest optimized feed diversity and quality together, evolving from DPP to SSD and soft spacing.
Building an MCP Ecosystem at Pinterest
Pinterest runs agent automation with internal MCP servers and a registry.
Unified Context-Intent Embeddings for Scalable Text-to-SQL
Embedding past SQL with intent and structure produces more accurate Text-to-SQL.
Pinterest Unifies Ads Engagement Modeling Across Multiple Surfaces
By combining HF and SR into a single model with surface-specific calibration and lightweight optimizations, Pinterest achieved better results.
Closing the Gap: Diagnosing the Online–Offline Discrepancy in Pinterest's L1 Conversion Model
Even when offline performance was strong, online CPA didn't improve—the causes were traced to features, embeddings, and the funnel.
Piqama: Pinterest's Quota Management Ecosystem
Piqama unifies quota lifecycle and right-sizing across Pinterest into a single platform.
How Pinterest drastically reduced Apache Spark OOM errors
Pinterest reduced Spark OOM failures by 96% with Auto Memory Retries.
Lightweight Ad Engagement Prediction with Two-Tower Models Served on GPU
GPU serving based on MMOE-DCN improved both accuracy and latency for ad engagement prediction.
Pinterest's Next-Generation DB Ingestion Technology
Pinterest built a unified CDC-based DB ingestion framework to reduce data latency and cut costs.
Beyond Two-Tower: Redesigning the Serving Stack for Next-Generation Lightweight Ad Ranking Models
To increase the expressiveness of ad ranking models, the team moved away from the existing Two-Tower architecture and built a new GPU-based serving architecture.
Ad Candidate Generation Using Behavior Sequence Modeling
This covers a technical methodology for generating ad candidates by modeling users' behavior sequences.