ML261002RetrievalMid-Senior1 positionRemote
Applied ML Engineer, Retrieval & Evaluation
Own retrieval quality as an engineering problem, from chunking and hybrid search through re-ranking and grounding, plus the measurement that proves any of it worked.
Quote ML261002
01What you would work on
- Build ingestion pipelines that survive the real corpus rather than the demo one
- Design chunking that follows the structure a document already has instead of counting characters
- Combine vector and lexical retrieval, re-rank the union, and measure what each stage contributed
- Set abstention thresholds so a system can answer “I do not have anything on that”
- Build evaluation harnesses that score retrieval separately from generation, because the fixes are different
02What we are looking for
- Solid Python, and comfort with embeddings, vector stores and classical information retrieval
- You can explain precision and recall at k to a client without jargon
- Experience tuning a retrieval system against a measured set rather than impressions
- Healthy scepticism about any benchmark number that did not come from your own corpus
03Nice to have
- Cross-encoder re-ranking in production
- Document intelligence: OCR, layout analysis, table extraction