Speakers: Ben Lackey (Neo4j), Ezra Uzosike (Google).

Building AI agents powered by knowledge graphs, working against real datasets and deploying Neo4j alongside Google Cloud Gemini Enterprise in a live cloud environment. The advertised path runs end to end: parse and load data, then construct agents that query and reason over the connected result.

The four stated outcomes, verbatim:

Build AI agents that leverage knowledge graphs
Integrate Neo4j with Google Cloud Gemini Enterprise in a production-ready
  environment
Traditional vs. AI-driven data parsing and loading
Practical approaches to graph-based AI using real-world datasets

Audience is given as developers, data engineers, data scientists and tech leaders.

The third bullet is the one with a live question behind it. Traditional versus AI-driven parsing and loading is a comparison, and comparisons are where the interesting failure lives: the two paths produce different graphs from the same source, and the workshop format means both get built in the same room against the same data. What to watch for is whether the session states an acceptance criterion for the AI-driven path, or only demonstrates that it runs. An extraction pipeline that has never been shown a document it should refuse is not yet a pipeline.

This repository already runs a code knowledge graph — GitNexus, currently 1473 symbols and 3469 relationships over this site's sources, used for impact analysis before edits. That graph is built by a deterministic parser. The question this workshop is positioned to answer is what an LLM-driven extractor buys over that, and what it costs in precision, on a corpus where the ground truth is known.

Related, and close enough to read together:

• DevFest:Extended Boston 2026 — "escape naive loops with graph engineering", 33 days later at the Google Boston office. Same thread, different vendor framing: DevFest treats the graph as agent control flow, this one treats it as the data substrate.
• Code Search and Code Graph MCP Servers on FreeBSD — where Neo4j already appears in this corpus, surveyed rather than run.
• Agent Memory as Institutional Knowledge — the retention side of the same question.

• What is the acceptance test for an LLM-extracted graph? Precision and recall against a hand-labelled subset, or eyeballing the visualization?
• Where does Gemini Enterprise sit relative to the graph — generating Cypher, or consuming retrieved subgraphs as context?
• What happens to the graph when the source changes? Incremental update, or rebuild?
• Does the "production-ready environment" claim survive contact with authentication and multi-tenancy, or is that out of scope for the session?