Hanna Lichtenberg is an AI engineer leading agentic search at Mixedbread, where she builds systems that help language-model agents find the evidence needed to answer difficult questions. Her work treats unreliable retrieval as a central obstacle to using capable models for research, finance, legal analysis, and other document-intensive tasks.
Lichtenberg studied predictive uncertainty and active learning at Otto von Guericke University Magdeburg, completing a 2023 bachelor’s thesis on uncertainty estimation that compared neural-network confidence measures and their usefulness for efficient data labeling. She subsequently studied at Technische Universität Berlin and worked with Starwit Technologies on identifying unusual or dangerous traffic situations. Her 2025 work on traffic anomaly detection used probabilistic forecasting and object tracking to predict vehicle movements, account for surrounding traffic, and assess how uncertainty and detection errors affect reliability.
At Mixedbread, Lichtenberg applies that emphasis on dependable evidence to retrieval-intensive knowledge agents. Her approach centers on four ideas:
- Closing the retrieval oracle gap: A model can reason effectively when supplied with the right documents yet fail when it must locate them itself. Mixedbread’s retrieval evaluations examine that disparity on demanding benchmarks including BrowseComp-Plus and OfficeQA-Pro.
- Matching search tools to evidence: Semantic search supports conceptual questions; metadata filters narrow the corpus; grep-style lexical matching finds exact phrases or identifiers. Lichtenberg designs agents to describe the information they need in natural language instead of defaulting to ineffective strings of keywords.
- Evidence-first parallel search: Her retrieval architecture divides complex questions into separate evidence needs, searches multiple avenues simultaneously, removes duplicate results, and ranks relevant documents while limiting search rounds, latency, and context consumption.
- Training the search trajectory: Smaller retrieval agents learn from larger-model examples through supervised fine-tuning and on-policy reinforcement learning. Their rewards assess both the final documents and the search process itself, including tool selection, query quality, exploration, and efficiency.
Lichtenberg described the design and training of these systems in a joint AI Engineer World’s Fair presentation with Mixedbread cofounder Aamir Shakir.
