Organizations increasingly rely on AI to process large volumes of free-form text, but reliably mapping that input to fixed, approved categories remains a major challenge—especially when queries are vague, novel, or inconsistent. Standard large language model (LLM) approaches often introduce hallucinated outputs or drift, limiting trust and real-world deployment. This session explores a production-ready architecture developed at Pinterest that combines LLM-based semantic augmentation with vector embedding and strict matching in a curated knowledge space. The system enables robust, explainable classification without hallucinations and supports scalable campaign automation. You’ll learn how this was applied to classify over 100,000 search queries into granular ad groups and generate tens of thousands of brand-safe ad creatives, driving performance without compromising control.

What You’ll Learn

• How to use LLMs for semantic enrichment of inputs and categories in a classification pipeline
• How to build a robust vector-embedding architecture that ensures consistency and explainability
• How to apply zero-shot classification at scale without hallucinations or drift
• How LLMs can be safely integrated to generate brand-aligned, large-scale marketing content 

Key Takeaways

• A reusable system architecture for safe and explainable zero-shot classification
• Insights on applying vector databases to anchor LLM outputs to curated knowledge sets
• Strategies for using LLMs in high-trust environments like ad tech and content moderation
• A look at how to scale generative workflows while preserving control and brand integrity

Who Should Attend
Machine learning engineers, applied researchers, data scientists, product managers, and technical marketers working in:

• AI Infrastructure and Platform Engineering
• Advertising Technology and Campaign Optimization
• Content Classification and Moderation
• Scalable Generative AI Systems
• Search, Recommendations, and Taxonomy Development