All work
Retrieval
Lumen Maps
2025
Multilingual semantic search across two million small-business listings, rebuilt from brittle keyword matching.
Scroll to resolve
+27%
Search-to-visit rate
Client
Lumen Maps
Role
Staff engineer, search
Stack
Multilingual embeddings · Vespa · Rerankers
Year
2025
The problem
People searched for “late noodles near me” in six languages, and keyword search returned laundromats.
What I built
I replaced it with a hybrid retriever — dense multilingual embeddings plus the old keyword index — and trained a small reranker on real click-through data.
The result
Search-to-visit rate rose 27%, and the long tail of non-English queries finally returned useful results.