The company builds person and company search for AI agents, covering over 1B people.
The entity resolution team merges what the company indexes into one canonical record per person or company, with evidence attached. Sources include social networks, job boards, web directories such as school and institution pages, and proof-of-work platforms. AI agents query this data directly.
These sources share no common identifier for people or companies. The work covers how to pick which records get compared, which pairs match, and how matches form entities, and those methods go into production.
Responsibilities
Build candidate generation with blocking and similarity search (LSH, sketching, learned embeddings) to surface likely matches among billions of records, and measure how much recall is lost.
Create matching models that judge whether two records refer to the same person or company, with calibrated confidence even when names, employers, schools, handles or dates conflict.
Cluster pairwise scores into one record per entity, and manage merges and splits as new data arrives.
Decide which source to trust for each field by weighing reliability, recency and corroboration, and keep provenance for every value.
Design evals with labeled sets and metrics for matching and merging when ground truth is partial, plus lineage that traces an error to the source or stage that caused it.
Example projects: generate candidate pairs across 1B+ person records with recall you can measure and defend; link one person's profiles across a social network, job board, school directory and proof-of-work platform when no field matches exactly; build an eval set that estimates person-matching error without labeling everything; resolve each company to one record across domains, name variants, subsidiaries and acquisitions.
Requirements
A PhD or research master's in CS, ML, data management or a related field, completed or finishing in 2026, with a thesis or first-author papers on similarity search and hashing, entity resolution or record linkage, information extraction, knowledge graphs, data integration or large-scale ML.
Strong Python and a systems language, with code others can read, such as a released paper repo or an open-source project.
Wants methods running in production on a billion records.
Cares about correctness and measures it.
Bonus: LSH, MinHash, sketching or approximate nearest neighbor search at scale.
Bonus: weak supervision, active learning or LLM-based labeling for matching.
Bonus: Spark, Ray or another distributed data framework.
Bonus: time at a startup or on a small team.
Skills: Python, Machine Learning, Natural Language Processing.