Aamir Shakir is the co-founder and chief executive of Mixedbread, a search company building retrieval systems that help language-model agents find relevant evidence across unstructured information. His central concern is that improvements in model reasoning mean little when search cannot locate the documents needed to answer accurately.
After studying at École polytechnique fédérale de Lausanne, Shakir co-founded Mixedbread in 2023. He co-authored the 2024 release of mxbai-embed-large-v1, an openly licensed embedding model, and BMX, a retrieval method combining keyword matching with entropy-weighted similarity and semantic signals. Mixedbread’s open-source Baguetter search library incorporates lexical, dense, and hybrid retrieval.
His subsequent work expanded from embedding models to failures throughout the search pipeline. Research on the OCR performance ceiling shows how transcription errors in scanned or visually complex documents damage retrieval and generated answers; multimodal document retrieval offers an alternative by searching visual information directly. He has also advocated open retrieval models and smaller systems capable of operating across cloud infrastructure and personal devices.
- Closing the retrieval oracle gap. At AI Engineer World’s Fair 2026, Shakir identified the difference between a model’s performance with known relevant documents and its weaker results searching noisy collections. He argued that agents trained around coding tools and keyword-heavy web searches often fail to exploit richer semantic retrieval; co-presenter Hanna Lichtenberg explained Mixedbread’s agent architecture and training.
- Making multi-vector search practical. His work on asymmetric quantization addresses the storage costs of fine-grained, late-interaction document representations, a constraint on deploying multimodal search at scale.
- Training stronger compact rerankers. His co-authored ProRank research investigates reinforcement-learning methods for improving how smaller language models rank retrieved documents.
