Most graph pipelines still make their judgment calls with uncalibrated scores. Drawing on 58 reproducible experiments, Jim Thornton will show how to go from a hand-tuned if score > 0.85: merge to answering those calls (should these two nodes merge, should this extracted fact be written, is this edge still true, which neighbor leads to the goal) with small typed decision models: models that return a probability for a fixed set of options in one forward pass.
You will see the pattern behind the wins: Cypher proposes candidates, the model decides, and the probability is stored on the edge so you can set review bands, audit and roll back. Jim will show the Cypher, the exact questions asked and the results: duplicate CRM companies decided correctly 96 times in 100 versus 62 for merging on name, 89 of 100 junk writes from an LLM extractor caught versus 5 for a schema check, and graph walks that reach 18 of 40 targets versus 4 to 7 for an LLM writing Cypher at each hop, for about 4 cents per 1,000 decisions.
Just as useful are the ideas that failed, and Jim will cover those too, along with when a plain classifier is enough and an exact, server-free layer that answers lookups straight from Neo4j's store files. You will leave with a strong sense of how best practices for entity resolution, ingestion, graph maintenance and GraphRAG are quickly changing with experiments you can rerun.
Speaker
Jim Thornton
AI Knowledge Engineering at Inbound Found
Jim Thornton is a content strategist at a digital agency specialized in growth engineering with AI. He has consulted for small startups and multi-billion dollar brands across industries. He lives in Philadelphia with his wife, daughter, and son. As a newish father, he looks forward to early bed times.
