The method uses a repeatable cycle of proposing, testing, and learning to fix agent behaviors without breaking existing functionality.
Pinterest solved embedding overfitting by delaying table updates until the final epoch, enabling multi-epoch fine-tuning that improved online saves by 1.09%.
The team fused metrics computation with Ray inference to eliminate data serialization overhead, reducing end-to-end wall-clock time by 25–32% across dataset sizes.
The tool captures runtime data to reconstruct modules locally in seconds, supporting both human debugging and autonomous AI agent workflows.
Pinterest has transitioned production inference to a fine-tuned 4B LLM for user journeys, retaining the legacy clustering system as a fallback, with initial tests showing a 1.1% lift in email click-through rates.
The platform reduces metric creation time from weeks to 1–2 hours and supports over 98% of experimentation metrics.
The mechanism uses Iceberg snapshot summaries to publish non-regressing watermarks, allowing downstream jobs to safely consume data from streaming pipelines.
Pinterest has released its 3-Tower Engagement Co-Train model, achieving a 1% CTR lift and a P99 latency of 30ms.
Pinterest reduced serving costs by up to 30% by introducing SSD serving and quantization technologies to the Manas platform.
Pinterest reduced embedding search costs by up to 30% by introducing quantization and SSD serving to the Manas platform.
Pinterest built a VLM serving stack using NVIDIA Dynamo and Blackwell GPUs, improving TTFT by 85x.
AI teams require a fundamental shift in operating models rather than simple tool adoption, necessitating that engineers evolve into strategists and managers into vision leaders.
Pinterest expanded Conditional Learning Retrieval (CLR), which reflects multiple intents, to the home feed to improve efficiency.
Pinterest grew Weekly Active Users (WAU) through use-case representation techniques that more precisely capture user intent.
Pinterest has revealed the design of RPP, its in-house CI/CD engine that securely manages hundreds of Terraform workspaces.
Pinterest solved performance degradation issues in distributed training to achieve near-linear training scalability for its foundation model.
Pinterest built a next-generation DB Ingestion Framework that automatically handles schema changes to improve the efficiency of its data pipelines.
Pinterest introduced a 'One Definition, Many Runtimes' approach to improve the cost efficiency and consistency of its user sequence data platform.
A test harness quantified agent skill invocation rates, and Codex improved more than Claude.
Pinterest improved ad relevance and ROAS with a sequential recommendation model that integrates real-time context.
Pinterest reduced network bottlenecks in root-leaf ML serving by trimming unnecessary features.
Pinterest redesigned its conversion-focused candidate generation model, boosting RoAS and conversion performance.
Pinterest improved content deduplication accuracy by refining URL normalization with MIQPS.
The cause of Ray training failures wasn't the GPU, but CPU starvation affecting the ENA driver.
AI Briefing