Doug Turnbull. Search relevance engineer who writes at softwaredoug.com about LLM-native search patterns, embedding techniques, and practical machine learning for information retrieval. Turnbull focuses on the mechanics of how language models interact with large classification vocabularies and retrieval systems, offering engineering-first takes grounded in production experience.
Operating themes
- Two-stage LLM pipelines Decoupling unconstrained generation from deterministic matching outperforms single-step constrained classification in large vocabularies.
- Embeddings as resolution layer Treating embeddings as a semantic resolution step rather than a primary retrieval mechanism reduces token cost and schema complexity.
- Intentional hallucination Using LLMs to generate plausible but fictional entities, then grounding them via embeddings, is a design pattern, not a failure mode.