Stephen Chin of Neo4j and Neeraj Deshmukh of Blitzy explain their graph-based approach to AI-assisted mainframe modernization in a 20-minute AI Engineer talk.
A formula sits inside a larger contract
- Their AWS CardDemo example connects interest calculation to batch jobs, record layouts, account updates, and transaction output.
- Translating the formula alone can lose behavior elsewhere in the system.
- A dependency graph makes relationships explicit so an agent can follow a change across multiple connected components.
Give each agent a traceable slice
- Graph relationships retain repository revision, touched symbols, source location, and extraction evidence.
- Updates on each merge keep the graph aligned with changing code.
- Each task packet combines the source slice, constraints, supporting evidence, and unresolved questions.
- The speakers’ “infinite code context” means retrieving the right material as needed, rather than loading the whole estate into one prompt.
Approve a plan, then verify behavior
- A human approves the transformation plan and its preservation requirements before code generation.
- Blitzy describes continuous testing, security controls, and thousands of agents sharing that plan.
- Compare old and new systems on outputs, persisted state, transaction records, errors, and complete workflows.
- Their one-cent interest discrepancy illustrates why a successful compile leaves financial parity unproven.
Keep the case-study scopes separate
- Deshmukh reports reverse-engineering a 35-million-line mainframe in four days; that claim concerns mapping, not completing its migration.
- A separate bank project reportedly took 18 days and added more than 4,000 tests against an estimate of 18 engineer-weeks.
- Another insurer example carries claims of nine-times-faster delivery and substantial autonomous code completion.
- These are vendor-presented cases, with no shared baseline artifacts or independent reproduction in the talk.
“You have to prove the parity.” — Neeraj Deshmukh, 12:03