The Knowledge Platform team builds Exa's knowledge engine. We continuously synthesize information from across the web and beyond into one of the world's largest open-domain knowledge graphs, with hundreds of billions of evidence-backed claims. The work has two halves: construction pipelines that decide what to trust, and retrieval engines that decide how best to search it. Our goal is to give agents answers to any factual question about the world with unprecedented precision and completeness.

In this role, you'll build the knowledge construction pipelines: turning everything from a 100k docs/s web crawl to vendor datasets into a graph of what is true about the world, and keeping it correct as data changes.

This is greenfield work with a massive scope. There is no off-the-shelf system for building a model of the world from the open web, so you'll make foundational decisions, from how claims are represented and resolved to how trust is assigned and how the graph evolves over time, that everything built on top of it depends on.

What you'll do

• Continuously synthesize the live web crawl, vendor deliveries and APIs, and other off-the-web sources into one typed graph: extraction that turns pages and records into ontology-shaped claims, each carrying its provenance, timestamp and confidence.
• Bring every new source in cleanly: versioned, append-only history with schema evolution, late-arriving and corrected records, and replayable cursors, so every claim traces back to the exact delivery that produced it.
• Decide what to trust: entity resolution across billions of records (blocking, pairwise scoring, clustering) into canonical people, companies, places and publications, then truth discovery that weighs source reliability, recency and corroboration to reconcile conflicting claims.
• Keep that picture current and convergent: streams arrive continuously, out of order and at very different rates, so the graph is eventually consistent by design. You'll build incremental computation over change data feeds, where each update is a diff of what changed and every job is idempotent and resumable, so the graph converges to the same state however runs interleave, fail or replay, and readers only ever see committed, consistent versions.
• Measure precision and completeness: labeled evals for extraction and resolution, per-source coverage and correctness metrics, and lineage that pinpoints which source or stage broke and why.

What we're looking for

• You've built and operated large-scale distributed data pipelines in production over petabytes of structured and unstructured data, with Spark over a lakehouse format (Delta Lake, Iceberg, Hudi).