Careers
We're small, and the problems are the good kind.
Distributed systems, retrieval quality, and a habit of publishing results that do not flatter us.
Open roles
Nothing posted right now.
Which is not the same as not hiring. If you read the research page and had opinions about it, that is a better introduction than a cover letter.
What it's like
Four things that are true.
The interesting problems are not the model calls
Entity resolution that holds under concurrent writers. Provenance bookkeeping that retracts a fact only when its last supporter drops. An ontology apply that is reversible because it filters rather than deletes. Orchestration that does not silently lose a document. The language-model part is the easy part.
Measurement over intuition
We ran 2,160 retrieval configurations to find out that the most intuitive one — weighting results by how many documents corroborate a fact — was the worst. That is the normal way decisions get made here, and it means being wrong in public reasonably often.
We publish the unflattering result
Our own research page says the knowledge graph has not been shown to improve answer accuracy over plain retrieval on our question sets. That is on the marketing site, not buried in a doc. If that sounds uncomfortable rather than appealing, this is probably not the right team.
Architecture rules are enforced, not reviewed
A custom lint plugin gates the contract-first structure in CI, with no suppressions permitted. Fewer arguments in code review, and the constraint is the same for everyone including whoever wrote it.