AI-Powered Search & Discovery

User search behavior had evolved from simple keywords to incorporating natural language, misspellings, and conversational phrases. We saw opportunities to improve semantic relevance, enhance understanding of user intent, increase flexibility for new offer types, boost discoverability of key promotions, and accelerate iteration and optimization cycles.

Starting from a winning hackathon proof-of-concept, I led the transformation into a production system using Databricks Vector Search. We built a comprehensive 50-metric evaluation framework with LLM-as-a-judge patterns, systematically tested 30+ iterations, and ultimately fine-tuned our own embedding model. The solution included Query Transformation for enriching vague queries and Multi-Search for fanning out generic terms into parallel searches.

The production system delivered exceptional results: 14.8% more offer unlocks, 6% increase in engaged users, 15% increase in engagement on bonuses, 72.6% decrease in searches with zero results, 60.9% fewer users encountering no results, and 60% lower search latency. These improvements translated directly to better user experience and increased revenue.