Bilge Yücel

Bilge Yücel is a senior developer relations engineer at deepset and a contributor to Haystack, the open-source framework for building search, retrieval-augmented generation, and AI agents. She develops practical approaches to multimodal retrieval, agent safety, and sovereign AI: systems whose data, models, infrastructure, and operations remain under meaningful organizational control.

Yücel worked as a web developer at Hipo before earning a master’s degree in artificial intelligence at KU Leuven, where she focused on natural language processing. She joined deepset as a developer-advocacy intern and became a full-time developer advocate, combining software development with tutorials, live coding, open-source community support, and technical education. She later advanced to senior developer relations engineer.

• Sovereignty across the entire AI stack. Yücel distinguishes four operational concerns: where data is processed, whether models can be replaced, who controls computing infrastructure, and whether systems remain observable and auditable. Her AI Engineer Europe session identifies concrete failure modes: embedding APIs that move sensitive information across jurisdictions, model-provider lock-in, unexpected CPU/GPU and Kubernetes requirements, and incident response dependent on outside vendors. She treats sovereignty as a spectrum calibrated to organizational risk.

• Model-agnostic orchestration and accountable agents. Her Haystack architectures combine explicit data flows, interchangeable models, versioned YAML pipelines, OpenTelemetry-compatible tracing, input and output guardrails, and human approval for consequential actions. She uses locally hosted Model Context Protocol tools and BM25-based tool discovery to expose relevant capabilities without overwhelming an agent’s context window.

• Retrieval complexity that earns its keep. Yücel compares agent-driven retrieval with conventional pipelines using latency, accuracy, robustness, and traceability; autonomous search becomes useful when requests are ambiguous, tasks require multiple steps, or available tools change dynamically. An Intel agent-building guide demonstrates her practical use of document retrieval, web search, external APIs, and open models.

• Multimodal retrieval across a shared vector space. Writing with Stefano Fiorucci, she shows how text, images, video, audio, and PDFs can support cross-modal search. Her multimodal agent workshop extends those ideas into agents with image understanding, document retrieval, conversational memory, and human-in-the-loop controls.