← All speakers

Zach Blumenfeld

Conference affiliation: Neo4j · 2026

Zach Blumenfeld is an AI Research Engineer at Neo4j working on graph representations of the information agents retrieve and the procedures they follow. His examples make the distinction concrete: a search result cannot establish how many employees know Python, and a correct loan decision does not establish that an agent followed lending policy. Zach uses connected records, decision traces and runnable procedures to make those questions answerable.

On this page

From matching records to guiding agents

Zach’s earlier technical writing tackled identity matching, recommendations and supply chains. His 2021 work on supervised entity resolution represented potential matches as relationships between records, preserving uncertainty rather than immediately collapsing them into one identity. Subsequent examples explored personalized news recommendations and supply-chain analysis.

At Neo4j, Zach worked as a graph data science product specialist and in technical marketing before moving into research engineering in 2026. He described the move as a return to his AI and machine-learning roots. His teaching connects that earlier graph analytics work with generative applications.

In a 2024 retail workshop, supported by colleagues helping participants, Zach demonstrated a fashion assistant built around clothing descriptions and purchase histories. Two customers could search for denim jeans and receive different recommendations: semantic search found relevant clothing, while shared purchasing patterns changed its ranking. Graph embeddings supplied additional products to pair with the results, and a language model assembled the selections into an email.

Zach also acknowledged where the demonstration needed further work. Multiplying semantic and purchase scores was a starting point, not an established ranking method; production recommendations needed labeled outcomes and comparisons between scoring methods. Restricting retrieval to a customer’s purchasing neighborhood could also exclude useful products when that history was sparse. His example retained semantic-search results even when the graph supplied no purchasing signal.

By 2025, Zach’s examples increasingly treated the data model as something an agent could inspect to choose and compose retrieval operations. His writing includes introductions to Neo4j’s GraphRAG Python package and Aura Agent. His independent experimental project, graph-nd, explores graph-grounded retrieval and agent workflows.

Matching retrieval to the question

Zach’s employee-skills example starts with an agent that reports five Python developers because its document-search tool returns five résumés. He extracts people, skills and accomplishments into a graph so the agent can count matching people instead of treating a retrieval limit as a workforce total. The same representation supports explicit skill-overlap calculations and collaboration queries. Adding internal project records expands an assumption inherited from individual résumés: an accomplishment can have several collaborators.

His introductory GraphRAG workshop develops several choices behind that approach:

  • Make the schema readable. Relationships such as “person knows skill,” supplemented with annotations and query examples, help a model interpret the data. Zach favors relatively small schemas for generated queries and specialist tools with defined query logic for complicated traversals.
  • Distinguish exact overlap from related experience. Semantic search can find related skills; graph traversal identifies the people who possess them. Combining the two lets a developer adjust how much a similarity calculation values shared skills versus related ones. Leiden community detection serves a different question: identifying groups with similar skill profiles across the workforce.
  • Account for the cost of derived relationships. Storing a skill-overlap relationship can avoid repeating a traversal, but it must be recalculated when the underlying data changes. Keeping vector search in the graph database can reduce synchronization work and potentially latency; an existing external vector store may still make sense on cost and performance grounds.

Use the structure already in the data

In the lakehouse workshop with Ben Squire and Ryan Knight, Zach demonstrates how graphs can guide retrieval without requiring an organization to move all its operational records. A metadata graph describes warehouse tables, columns and join paths; the agent uses it to compose SQL while the records remain in BigQuery. He distinguishes this from importing records for fast recursive traversals or graph algorithms, where storing the data itself in a graph can be valuable. Continuous synchronization and restrictions on moving sensitive data also affect that choice.

For documents, Zach starts with headings, section order and existing cross-links. A deterministic loader turns those into a navigable graph without requiring an LLM to extract every entity. Search helps locate an entry point, document trees supply surrounding context, and community detection groups interlinked material. In the workshop’s auto-repair scenario, those views help compare repair histories with manuals and bulletins to find recurring field problems that lack documentation.

This reflects a change in Zach’s practice: he says stronger agents sometimes let him omit explicit entity nodes he previously would have required. Precise domain ontologies still matter where the application needs them. He also keeps the demonstration’s limits visible: the document and warehouse graphs are separate, with the agent connecting them through extracted codes, and he had not benchmarked the workshop patterns at larger scales. Navigation complements search; its usefulness still depends on document names, links and content remaining accurate.

From remembering decisions to running procedures

Zach’s decision-trace work extends retrieval to precedents. His financial-agent example connects customer records and policies with previous decisions, combining semantic similarity with graph embeddings that represent the structure of decision traces. He distinguishes conversation history, persistent entity knowledge and reasoning memory: the recorded steps and relationships behind earlier decisions.

A stored trace is not automatically a good precedent. Zach identifies open questions about weighting recent events, assessing decision quality and deciding which traces to retain. He credits colleague William Lyon with the scaffolding tools he demonstrates, including tools for getting started with context graphs.

His subsequent graph-engineering presentation examines how successful runs can become reusable procedures. He distinguishes the graph that executes a workflow, the trace of what happened in one run and the procedure specifying which paths a task permits. A trace explains a route after the fact; a runnable procedure supplies routing rules for future runs.

In the team’s lending experiment, scripts apply hard policy rules first, automatically approving or denying applications those rules settle. Only unresolved cases reach a decision model. A separate route handles human escalation, and generated reports pass through a checking loop. The procedure retains links to the runs and records from which it was distilled, allowing reviewers to inspect its origins and revise its logic.

Storing the permitted routes alongside actual execution histories makes a specific distinction visible: whether an approval followed a policy rule or required model judgment. It also exposes branches that the procedure allows but no recorded application has exercised.

Read the topics behind these talks

6 conference talks

Loading more talks and highlights…

References