AI Engineer
← All speakers

Related work and projects

Bio, Work & Ideas

Zach Blumenfeld

Conference affiliation: Neo4j · 2026

Zach Blumenfeld is an AI Research Engineer at Neo4j developing graph-based approaches that help AI agents retrieve evidence, understand relationships and reason from previous decisions. His work applies graph data science to problems ranging from fraud detection and identity matching to agentic GraphRAG and decision-aware AI systems.

From graph analytics to AI agents

Blumenfeld joined Neo4j as a graph data science product specialist and subsequently worked in technical marketing before moving into research engineering. In 2021, he applied supervised entity resolution to identity matching, representing probable matches as graph relationships instead of prematurely merging uncertain records. His subsequent projects explored personalized news recommendations and supply-chain analysis.

As generative AI matured, he brought those techniques into retrieval and agent development, writing about Neo4j’s GraphRAG Python package and Aura Agent. He also created graph-nd, an independent experimental project for graph-grounded retrieval and agent workflows.

  • Retrieval must answer the actual question. Blumenfeld illustrates the limitations of document search with a workforce assistant that retrieves five résumés and mistakenly concludes a company has five Python developers. His agentic GraphRAG demonstration replaces that guesswork with a graph connecting employees, skills and accomplishments, enabling agents to count developers, compare expertise and identify collaborators through inspectable Cypher queries.
  • Hybrid graph-and-vector retrieval makes similarity actionable. Semantic search can identify related skills; graph traversal establishes who possesses them and how individuals overlap. Blumenfeld combines these techniques with Leiden community detection and specialized retrieval tools while acknowledging practical tradeoffs: derived relationships require maintenance as data changes, and consolidating vector search inside a graph database affects cost, synchronization and latency. His GraphRAG workshop develops these patterns hands-on.
  • Graph-shaped knowledge layers can leave operational data in place. Rather than copying entire warehouses into another database, Blumenfeld uses graph representations of table relationships to guide agents toward valid joins while source records remain in existing systems. His lakehouse workshop also applies navigable document hierarchies and community detection to questions such as which recurring operational problems lack documentation.
  • Decision traces and context graphs give agents precedents. Blumenfeld distinguishes conversational memory, persistent knowledge about entities and reasoning memory that connects previous decisions with their circumstances and outcomes. His work on decision-aware agents combines semantic and graph-structural similarity to retrieve relevant precedents while treating temporal relevance, decision quality and automated retention as unresolved engineering questions.

Read the topics behind these talks

4 conference talks

▶ Watch ↗

AI Engineer World's Fair 202515:25

Agentic GraphRAG: Simplifying Retrieval Across Structured & Unstructured Data — Zach Blumenfeld

Read the full talk →

Key ideas

Scroll to read ↓

An employee assistant needs more than relevant resumes to count skills or find collaborators. A small knowledge graph gives its retrieval tools explicit relationships to query and extend.

  • Give the agent a model of the data
    0:00 ↗
  • Five retrieved resumes are not a workforce count
    2:00 ↗
  • Similarity, distributions, and collaboration need different operations
    3:56 ↗
  • Extract people, skills, and accomplishments
    5:40 ↗
  • Count and compare through explicit relationships
    8:32 ↗
  • Add tools for richer employee comparisons
    10:25 ↗
  • Extend individual accomplishments into shared work
    12:27 ↗
  • Find who delivered AI work together
    14:10 ↗
▶ Watch ↗

AI Engineer World's Fair 20261:59:10

AI on Your Lakehouse: Context Comes in Shapes, Not Queries

Read the full talk →

Key ideas

Scroll to read ↓

A repair copilot needs more than matching passages and valid SQL. Warehouse connections, navigable documents, and link communities give it different ways to assemble the context a question requires.

  • The repair history and the manual are only the beginning
    2:08 ↗
  • A graph, an agent, and a runnable workspace
    7:37 ↗
  • Connections: use a graph to choose the SQL
    21:57 ↗
  • When metadata is enough—and when to import records
    34:26 ↗
  • Table of Contents: preserve the structure documents already have
    43:40 ↗
  • Build an outline that can stop at the right depth
    52:48 ↗
  • Keep the vocabulary useful to the query author
    1:03:09 ↗
  • Search the text, then restrict it to a subtree
    1:08:47 ↗
  • Names, freshness, and references are part of retrieval quality
    1:14:38 ↗
  • Themes: discover groups through document links
    1:21:43 ↗
  • Community descriptions, changing data, and trust
    1:33:23 ↗
  • From a diagnostic code to a repair recommendation
    1:39:19 ↗
  • Find the mismatch between the library and the field
    1:45:31 ↗
  • Theme coverage and the join the agent still performs
    1:52:10 ↗
▶ Watch ↗

AI Engineer World's Fair 20251:18:35

Intro to GraphRAG

Read the full talk →

Key ideas

Scroll to read ↓

Build a skills graph in Neo4j, combine exact overlap with semantic similarity, extract people from biographies, and expose the resulting retrieval patterns to a LangGraph agent.

  • From employee data to an assistant
    0:00 ↗
  • Put domain logic in the graph
    9:05 ↗
  • Load people, skills, and stable identities
    19:34 ↗
  • Expand from Lucy to a local skills neighborhood
    26:43 ↗
  • Decide how much Cypher the model should write
    30:56 ↗
  • Count shared skills, then decide whether to store the result
    32:18 ↗
  • Use Leiden to make skill communities explicit
    35:52 ↗
  • Make the schema readable to an agent
    41:28 ↗
  • Embed descriptions, then make similarity traversable
    44:25 ↗
  • Combine exact evidence with semantic expansion
    49:27 ↗
  • Turn biographies into the same graph model
    55:17 ↗
  • Keep document structure when context crosses sections
    1:01:18 ↗
  • Package the graph patterns as four expert tools
    1:04:40 ↗
  • Let the agent choose and chain the tools
    1:11:11 ↗
  • Extend the assistant with annotated-schema queries
    1:15:40 ↗
▶ Watch ↗

AI Engineer Europe 202620:12

Why your agents need decision traces, not just documents — Zach Blumenfeld, Neo4j

Read the full talk →

Key ideas

Scroll to read ↓

A financial agent needs more than customer facts to justify a decision. Context graphs connect those facts to precedents, policies and outcomes—and expose the unfinished work of learning from them.

  • What does an agent need to make a decision?
    0:16 ↗
  • Following the financial agent’s retrieval
    4:04 ↗
  • Searching for a similar decision structure
    7:23 ↗
  • Starting with a generated application
    8:34 ↗
  • From conversation history to connected memory
    13:08 ↗
  • Exploring the system—and representing time
    14:56 ↗
  • Preparing a schema for existing information
    16:53 ↗
  • Who decides which traces become experience?
    18:41 ↗

References