Stephen Chin
Conference affiliation: Neo4j · 2026
Related work and projects
- cnc-star-creator ↗github.com · Source ↗
- JavaFX-Spring ↗github.com · Source ↗
- JFXPoetry ↗github.com · Source ↗
- LUGOD-Pi-Test ↗github.com · Source ↗
- MaryHadALittleLambda ↗github.com · Source ↗
- widgetfx ↗github.com · Source ↗
Stephen Chin is vice president of developer relations at Neo4j and coauthor, with Michael Hunger and Jesús Barrasa, of GraphRAG: The Definitive Guide. He teaches developers how to combine language models with connected data: use similarity search to find a starting point, then follow relationships to recover the dependencies, history, and decisions an answer requires.
On this page
From Java development to developer education
Stephen worked as a Java developer migrating COBOL applications off mainframes roughly three decades before his recent work on AI-assisted modernization. Those projects could take years and still miss consequential corner cases. He recalls his team making mistakes with decimal arithmetic and rounding—details that could leave a financial calculation looking plausible while producing the wrong result.
He subsequently built a career in Java advocacy, technical books, conference programming, and projects such as JavaFX-Spring, which integrated JavaFX applications with Spring technologies. At Oracle, he worked in Java advocacy and developer marketing, becoming director of developer marketing by 2018. His account of joining JFrog in 2019 made an economic argument for open source: successful ecosystems depend on supporting the companies and maintainers producing their underlying software.
At JFrog, Stephen became vice president of developer relations and served on the governing boards of the Continuous Delivery Foundation and Cloud Native Computing Foundation. His subsequent work at Neo4j centers on knowledge graphs and enterprise AI. He also serves as program chair and a board member of the LF AI & Data Foundation.
Developer education remains central to that work. Stephen promotes GraphAcademy as a free, hands-on route from learning Cypher queries to building applications with GraphRAG and agent memory. His collaborative book develops these subjects alongside agent architectures and industry use cases. In his work with Michael Hunger, Stephen explains the case for graph-grounded AI while Michael develops the implementation patterns; Michael’s team builds tools including Knowledge Graph Builder and Neo4j’s MCP servers.
Finding the connections an agent misses
Stephen’s central distinction is between similarity and relevance. An embedding can locate text resembling a question without recovering all the facts needed to answer it. His recommended starting architecture gives vectors and graphs complementary jobs: vector search identifies promising entry points, and graph traversal follows their relationships to assemble context for the model.
He also identifies where the extra machinery earns its place. A question answerable with one similarity search may need no graph. Graph retrieval becomes more useful when an answer depends on several connected facts. A request to update the presentation he last delivered with a colleague, for example, requires connecting two people to an event, its date, and the presentation used there. Matching the words in the request is only the entry point.
His demonstrations explore different consequences of that distinction:
- Security answers need affected versions and remedies. In his context-engineering demonstration, Stephen loaded supply-chain and vulnerability documents into a knowledge graph. A fixed retrieval pipeline located information about a Jackson-library vulnerability. An agent using Neo4j’s MCP tools went further: it inspected the graph schema, issued successive queries, and retrieved supporting text to assemble affected versions, attack details, and remediation advice. The fixed pipeline was faster; the iterative approach produced a more detailed answer.
- Persistent memory must support retrieval across sessions. For CrabRAG, Stephen built a digital twin of his home network and compared vector-backed and graph-backed retrieval from the same original information. He isolated the demonstration environment so the agents had to answer from memory. In that comparison, graph traversal connected internet exposure to outdated software and exposed management interfaces, while the vector-backed assistant returned incomplete information or instructions to investigate manually. Stephen said he subsequently patched the problems the demonstration uncovered.
- Decision history needs connections to business facts. His context-graph approach brings together short-term activity, persistent domain knowledge, and recorded decision traces. A financial-services demonstration connected customer history, an earlier rejection, policies, and risk factors to a lending recommendation. A person reviewing the recommendation could inspect the circumstances behind it and draw on previous decisions, rather than receiving an answer detached from its history.
For Stephen, inspectability also gives developers a way to improve retrieval. They can examine which nodes and relationships reached the model, adjust extraction, remove duplicate entities, or change traversal. That makes the graph useful both as context for the agent and as something an engineer can investigate when an answer goes wrong.
His discussion with Pfizer’s Jonathan Lowe extends these architectural questions into enterprise adoption. Jonathan described his team’s graph-based work on life-sciences knowledge transfer; Stephen pressed on the business case, the cost of replacing existing systems, and the organizational resistance a new architecture can encounter. The discussion connects retrieval quality to the practical question of whether an organization will fund and use the application.
Modernization requires more than a translated formula
Stephen’s mainframe-modernization session with Blitzy’s Neeraj Deshmukh returns to the kind of work he did as a Java developer. Neeraj used an interest-calculation example to show why translating a formula is insufficient: its behavior also depends on the batch job that calls it, the COBOL copybook defining record precision, and the account data it updates. He described how Blitzy’s code graph retains the file, source span, and repository revision behind a relationship so an engineer can inspect the dependency.
Stephen connected that example to his own experience with migration errors and to his concern that larger context windows do not ensure models attend to the relevant information. A translation can compile and produce a plausible number while changing rounding behavior by a cent. Across a financial system, preserving behavior requires recovering the surrounding rules and testing the result against them. Here, his argument for connected context has a concrete consequence: an agent needs to understand the system using the code before it changes the code.
Read the topics behind these talks
8 conference talks
▶ Watch ↗AI Engineer Summit 202520:59
Anchoring Enterprise GenAI with Knowledge Graphs
Explore the talk →
▶ Watch ↗AI Engineer Code 202526:50
Context Engineering: Connecting the Dots with Graphs — Stephen Chin, Neo4j
Explore the talk →
▶ Watch ↗AI Engineer World's Fair 202620:42
CrabRAG: Why Automated Assistants Need Graph Memory, Not More Tokens
Explore the talk →
▶ Watch ↗AI Engineer World's Fair 202519:46
Practical GraphRAG: Making LLMs Smarter with Knowledge Graphs
Explore the talk →Loading more talks and highlights…
References
- github.com/steveonjava/cnc-star-creator ↗
- github.com/steveonjava ↗
- github.com/steveonjava/JavaFX-Spring ↗
- github.com/steveonjava/JFXPoetry ↗
- github.com/steveonjava/LUGOD-Pi-Test ↗
- github.com/steveonjava/MaryHadALittleLambda ↗
- github.com/steveonjava/widgetfx ↗
- ai.engineer/talks/agentic-graphrag-ai-s-logical-edge ↗
- ai.engineer/talks/anchoring-enterprise-genai-with-knowledge-graphs ↗
- ai.engineer/europe/2026 ↗

