How We Mapped a 35-Million-Line Mainframe in 4 Days - Stephen Chin, Neo4j & Neeraj Deshmukh, Blitzy
Loading slides…
Read the talk
How Blitzy Mapped a 35-Million-Line Mainframe in Four Days, and What It Takes to Move the Money
Stephen Chin of Neo4j and Neeraj Deshmukh of Blitzy follow an interest payment from a COBOL batch job to its Java replacement, showing how to find the dependencies an agent needs and test whether the money comes out right.
From a talk by Stephen Chin and Neeraj Deshmukh
At a glance
Ideas worth remembering
Scope a financial migration around a complete posting. Identify the batch caller, record layout, persisted balance, transaction output, and failure outcomes before assigning the arithmetic.
Keep the repository revision and source span behind each dependency. Refresh the graph on merge so the next agent can work from the code that actually exists.
Use the existing system to establish expected behavior. Compare equivalent runs at each write and across the complete workflow, and ask a domain expert to decide which existing defects should change.
The mobile banking app has a nightly batch job underneath it
A customer can bank from a phone while decades-old software handles the account. Under that interface, business rules determine what should happen, nightly batch jobs do the work, and data contracts govern what one program can expect from another. Stephen Chin starts here because it explains why engineers hesitate over a one-line change. The line is easy to edit. Finding every process whose behavior depends on it can take much longer.
About 30 years earlier, Chin was a Java developer migrating COBOL applications off mainframes. These were multiyear projects. His team could translate the code and test it, yet still miss a corner case that surfaced as a small, consequential bug. In his experience, projects often failed. Neeraj Deshmukh takes up that testing problem with an example small enough to follow from the job invocation to the money it records.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Follow an interest posting through AWS CardDemo
AWS CardDemo is a public demonstration system developed by Amazon. Deshmukh picks one operation, computing interest. A batch job calls the interest program. The program works with record layouts defined in COBOL copybooks and produces VSAM data. To follow one posting, start with the inputs and the job that supplies them, inspect the layouts the program uses, then follow the account update and the outgoing transaction. The arithmetic sits partway along that path.
Now give an AI coding tool the instruction to convert the interest formula from COBOL to Java. It can find the expression and produce a translation. That instruction gives it little reason to inspect the precision of the record holding the amount, the update to the account, the transaction output, or the contract under which the batch job runs. Each affects what it means to post the right amount. Deshmukh describes the missing knowledge as the other 14 places the logic is used. A reviewer looking only at the translated expression has the same blind spot.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Three ways to retrieve context, with different results
Deshmukh asks why expanding context would fail to solve this. Chin distinguishes three retrieval approaches. A vector database finds material by similarity. It may surface another passage about interest, while a relevant layout matters because the program includes it. A relational database offers a larger repository of facts, accessible, for example, through an MCP server issuing SQL queries. Chin describes that pattern of use as a single-hop lookup. A graph supports multi-hop retrieval: begin at a program, find the job that executes it, inspect the copybook it includes, and follow its outputs. Each connection gives a reason to retrieve the next piece.
There is also a limit to what a model can use once the material has been found. Chin contrasts an advertised million-token window with his estimate of 200,000 or fewer usable tokens, citing declining accuracy as the window fills. The speakers also give a rough effective-context estimate of 20 to 30 percent and discuss codebases exceeding 200,000 to 300,000 lines. Those are different units and rough estimates. The practical problem is selection. A prompt presents material in sequence, while a program's caller, layout, and output may occupy distant parts of a repository.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Give each dependency a source location and a revision
In the graph of CardDemo, the interest program connects to the job that executes it, the copybooks it includes, and the outputs it produces. Deshmukh draws attention to what accompanies each edge. It retains its identity, its origin in the repository, the revision, and the symbols it touches. The evidence identifies the file, the span of source, and how the relationship was extracted. An engineer investigating a dependency can go back to the code that established it.
Blitzy updates the graph on every merge and invalidates relationships that have ceased to hold. That matters during a migration because the map is also an input to subsequent changes. Once a program or its dependencies change, an agent assigned to the next job needs the relationships supported by the merged revision. An accurate inventory of the original mainframe would, on its own, describe an increasingly old version of the system.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Package the evidence for one agent’s assignment
Each Blitzy agent receives a task context packet. It contains the task description, where the work belongs in the graph, the relevant slice of source, the constraints on the change, and the available evidence. The packet also identifies unanswered questions the agent is supposed to work on. In the interest example, the batch caller and the included layout belong together in the assignment even though they live outside the formula being translated.
This is what Deshmukh means by “infinite code context.” The software estate can be much larger than one prompt because an agent can access the information relevant to its task when it needs it. The graph supplies the connections. The packet narrows those connections to a particular decision and distinguishes what the source already establishes from what still needs investigation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Approve the transformation before generating code
Deshmukh likens Blitzy's process to compiling the knowledge in an enterprise. First it reverse-engineers the software estate to extract business logic, workflows, and hidden dependencies. The user then supplies requirements and rules for the proposed change. These feed an agent action plan, the specification that states the scope, what will be transformed, and which behaviors must carry over. A domain expert reviews that plan against the intended outcome and approves it before agents begin generation.
Agents have access to the repository and instructions for building the environment and running the software. When the source leaves a behavioral question open, they can execute the program, observe what happens, and use that observation while making the change. The deliverable has two parts. A pull request contains the tested code. A project guide documents what changed and the evidence used to validate it. That gives the person reviewing a consequential change both the implementation and an account of how it was checked.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Thousands of agents share a plan and test as they go
Deshmukh pairs what he calls hyperscale orchestration with extreme validation. Thousands of agents share the approved plan, while each works with the context relevant to its assignment. For a split such as interest calculation and account updating, the common specification matters: the work can be divided, but both pieces still have to satisfy the same posting behavior. Deshmukh says Blitzy spends more compute on testing than on code generation.
He identifies four controls throughout the process. System-level context exposes dependencies beyond the component being edited. Continuous validation checks the code during generation. Explicit guard layers apply security, governance, and compliance requirements, particularly important for financial systems. Finally, a human owns the decisions that matter. An existing defect is one such decision. If the replacement is supposed to reproduce today's behavior, correcting a defect changes the target and requires an explicit choice.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Compare what each operation returns, writes, and emits
For the interest operation, preservation has several observable dimensions. The monetary amount has a precision and a rounding behavior. The account has a balance after the update. A transaction record leaves the program. Failures have outcomes too. Blitzy runs tests against the existing system as a baseline, exercises the modern system with equivalent inputs, and compares the results operation by operation. A matching return value leaves open whether the program persisted the same balance or emitted the same transaction.
The checks accumulate. Generated code must compile and pass unit tests. At runtime it must integrate with the existing framework. Each write must achieve data parity, and the full workflow must reproduce the baseline, including error outcomes for the failure cases exercised. Deshmukh sets the target at zero differences for the operations being compared. He gives a customer's acceptance result as an example: at a Fortune 100 insurer, a transpiled job produced output identical byte for byte to its golden mainframe baseline. That is a concrete comparison for a particular job.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The Java calculation compiles and misses by one cent
Chin remembers making decimal-arithmetic and rounding mistakes as a Java developer. Did the AI do better? In Deshmukh's interest example, the generated Java compiled and produced an amount that looked reasonable to someone familiar with Java. On the test input, the Java calculation rounded up and posted one cent more than the COBOL baseline.
The mainframe run supplied the decisive expected value. A unit test written from the translated expression could have endorsed the Java result, and a reviewer inspecting the amount in isolation could easily have accepted it. Comparing the two implementations on the same operation exposed the cent. Deshmukh asks the audience to imagine that discrepancy recurring across a million transactions. Even at one cent, the team needs to know which accounts and records its calculation would change.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
An 18-day migration, a four-day map, and a policy platform
Deshmukh reports three engagements with different deliverables. At a tier-one bank, the starting point was a monolithic trade-posting engine with stored procedures and writes across databases. Blitzy converted it into six capability modules using PostgreSQL, Java, TypeScript, and Python. Deshmukh reports full behavioral parity, more than 4,000 tests, and a modern codebase less than half the size of the old one. Delivery took 18 elapsed days. The bank's original plan had allocated 18 engineer-weeks, a labor estimate rather than an elapsed-time measurement.
On the 35-million-line mainframe, the four-day result was reverse engineering. According to Deshmukh, the customer had already spent about eight months and several million dollars with a systems integrator and had no usable result to show. The customer described the earlier experience as inspecting individual modules in a dark room with a pencil light. Seeing the mapped system felt like switching on the overhead light. What the customer could now see was how the pieces related across the estate.
The third engagement involved modernizing a 25-million-line policy administration monolith. Deshmukh says Blitzy completed the work nine times faster than he estimates an engineer using IDE-based agentic tools would have done it. He also reports that 85 percent of the code was completed autonomously and that almost 90 percent of that portion made it into production.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Use the updated interest path to introduce a promotional rate
Deshmukh returns to CardDemo with a different assignment. Suppose the business now wants a promotional interest rate. The modernization has left the team with a graph of the upgraded posting path, established governance rules, and an expanded test suite. Engineers can trace where the new rate should apply and use the existing comparisons for postings that should continue to behave as before. The domain decision concerns which behavior to change. The earlier work supplies evidence about everything that decision touches.
Deshmukh extends that reasoning to new products, feature development, security detection and remediation, and evidence for compliance audits. Chin describes moving from code relationships to business rules and processes, with possible applications in tracking rule changes, examining DORA compliance across a system, and investigating financial fraud. For practitioners who want to inspect the example, Deshmukh points to the Blitzy Sandbox, which includes the AWS CardDemo modernization and accepts users' own code. Chin points to the Nodes Conference for 24 hours of sessions on building and applying knowledge graphs.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Full transcript
Loading transcript…