How We Mapped a 35-Million-Line Mainframe in 4 Days - Stephen Chin, Neo4j & Neeraj Deshmukh, Blitzy

Stephen Chin of Neo4j and Neeraj Deshmukh of Blitzy show why AI struggles to modernize the decades-old systems underneath banks and insurers, and how a knowledge graph fixes it. Using a COBOL interest calculation as the running example, they explain that AI can convert a formula but misses the batch jobs, data layouts and other places the logic is used, and that bigger context windows don't solve it. They walk through Blitzy's approach of mapping the whole codebase as a graph, giving thousands of agents the right slice of context, testing at every step, and proving the new code matches the old down to the cent, with results from a tier-one bank, a 35-million-line mainframe and a Fortune 100 insurer. Chapters: 0:00 Intro 0:14 The old systems under modern finance 1:58 Example: computing interest on a mainframe 3:06 Why AI misses the rest of the system 4:02 Why bigger context windows don't fix it 6:03 Turning code into a graph 7:08 Giving every agent the right context 7:53 Blitzy's process, with a human in the lead 9:32 Thousands of agents and constant testing 11:35 Proving the new code matches the old 13:28 The one-cent rounding bug 14:27 Results from banks and insurers 16:37 Beyond modernization 18:47 Resources Stephen Chin https://neo4j.com Neeraj Deshmukh https://blitzy.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!