Machine Learning Engineer, Retrieval

About the role

We keep finding the same thing across clients. Teams tune the model for weeks and leave the index alone, and the index was the constraint the whole time. This role exists because retrieval quality is the highest-leverage work in applied AI and almost nobody staffs it deliberately. You will own chunking, embedding, ranking, hybrid search and the measurement that proves any of it helped.

What you will do

• Build and tune retrieval systems across client corpora that were never designed to be searched.
• Run offline evaluations that predict production behaviour, and be honest when they do not.
• Work on ranking: hybrid keyword and vector, reranking, and knowing when a simpler approach wins.
• Instrument retrieval in production so a regression is visible before a client reports it.
• Publish what you learn internally, and occasionally on our blog.

What we need from you

• Practical experience with embeddings, vector search and the failure modes of both.
• You have measured a retrieval system rather than assuming it worked.
• Solid Python and SQL. You are comfortable in a database rather than only above it.
• An honest relationship with metrics: you know what recall at k does and does not tell you.

Useful but not required

• Information retrieval background, formal or self-taught.
• Experience with multilingual or heavily domain-specific corpora.
• You have built a reranker and can say whether it was worth it.