← All speakers

Andreas Kollegger

Conference affiliation: Neo4j · 2026

Andreas Kollegger, also known as ABK, is Director of Applied AI Research at Neo4j, working on knowledge graphs for retrieval and agent decision-making. His work connects two practical problems: finding information through relationships among records, and giving autonomous software enough organizational context to decide whether it should act.

On this page

Making connected data usable

Andreas’s earlier career included scientific-mission systems at NASA and medical informatics in Zambia, where he encountered graph databases. He subsequently joined Neo4j’s early core-engineering organization and contributed to the Neo4j 1.3 manual. His responsibilities expanded into product experience design, developer advocacy, and generative AI.

As product manager for Neo4j Bloom, Andreas addressed the difficulty of exploring connected data without writing database queries. His account of Bloom presents visualization as a tool for exploration, analysis, and communication: people can examine the relationships relevant to their work and share what they find.

His knowledge-graphs course applies connected data to retrieval-augmented generation, combining language models, Cypher queries, and structured relationships when an answer requires information scattered across documents and entities. His agentic knowledge-graph construction course extends that work to coordinated agents that propose schemas, extract relationships, and connect supplier records, product data, and customer reviews. The accompanying workshop gives developers a practical setting for multi-agent graph construction.

Let questions shape the graph

Andreas recommends starting with representative data and a simple schema, then testing whether the graph answers its intended questions. More entity labels and relationship types can make an organizational model complicated before they make it useful. In his contributions to a graph-analytics workshop, he recommends refining generic categories only when the data and evaluations show a need for greater specificity.

The same reasoning guides his choice between hierarchical summaries and knowledge-graph representations: begin with the questions people need answered, then choose a structure that supports them. Connecting passages lets a system investigate what they share, how they differ, and which communities they form—questions that retrieving isolated passages does not directly address.

Andreas also distinguishes a vector index from a complete search system. In the workshop, he explains how Neo4j’s property graph supplies filtering around vector retrieval, including predicates applied to the nodes returned by an index. Similarity identifies candidates; their properties and relationships help determine which candidates belong in the answer.

From retrieving context to deciding whether to act

In Context Graphs for Explainable, Decision-Aware AI Agents, presented with Zaid Zaim, Andreas develops a decision-making framework for situations an agent’s instructions have not anticipated. An assistant told to keep a refrigerator stocked with Red Bull might place an order while overlooking an upcoming rent payment. The missing information is a competing obligation that changes whether the purchase makes sense.

His framework makes several responsibilities explicit:

  • Frame the particular decision. Identify the objective, the events that led to it, and the environment in which the agent is operating. Compare prior decisions with current organizational rules; a precedent may cease to apply when those rules change.
  • Assess the consequences. Consider reversibility, the cost of being wrong, and whose circumstances matter. His medical example illustrates the danger of applying a generally successful treatment to someone in the subgroup for whom it could be harmful. The decision also needs an explicit objective: maximizing a benefit and minimizing a cost can produce different choices.
  • Separate proposals from permission. Andreas favors focused agents with distinct responsibilities. One develops alternatives and their advantages and disadvantages; another checks whether it has enough certainty and authority to act. It can execute a choice, escalate to another agent or a person, or defer action.

The process preserves decision precedents: alternatives considered, reasoning, decisions or deferrals, and actions taken. Future agents can consult that history while checking whether the present circumstances still justify the earlier choice. Andreas presents this as a general framework whose implementation must be specific to the domain.

His work on decision-making as constraint satisfaction also connects symbolic reasoning and language models to enterprise problems such as compliance, loan approvals, and resource allocation. His aie-matrix project, a virtual conference populated by autonomous personal agents, explores agents interacting in a shared setting. Across these projects, connected context includes both what an agent can learn and the rules, obligations, and prior decisions that should shape its next action.

Read the topics behind these talks

3 conference talks

Loading more talks and highlights…

References