About Neeraj Deshmukh
Neeraj Deshmukh leads engineering and research at Blitzy, an autonomous software development platform that uses thousands of coordinated agents to modernize and extend large enterprise codebases. Blitzy Research is the team behind the company's published SWE-Bench Pro results.
In an AI Engineer talk with Neo4j's Stephen Chin, he walks through Blitzy's method on the AWS CardDemo mainframe application: reverse-engineer the whole software estate into a knowledge graph, produce an agent action plan a domain expert approves, give each agent a task context packet, and prove parity between the old and new systems on every operation.
Notable Positions
On Why AI Fails at Legacy Code
The problem is not that AI can't write code; it's that the AI doesn't know the 14 other places a piece of logic is used. Blitzy's answer is "infinite code context": not remembering the whole estate but having the right information at the right time, drawn from a knowledge graph that keeps repo, revision, symbols and source span on every edge.
On Validation
Blitzy spends more compute on testing than on code generation. Parity between baseline and modern systems is tested on inputs, outputs, persisted state and error outcomes, and differences must be zero. A one-cent rounding difference between COBOL and Java is his standard example of a bug that compiles and looks correct.
On Human Control
Nothing happens without a human approving the plan, and defects become explicit decisions. He calls it "human in the lead." See Human-in-the-Loop.
Key Quotes
- "The failure here is not that AI can't write the code. It's that AI doesn't know the 14 other places where that logic, that code is being used."
- "We actually spend more computes on testing than we do on code gen."
- "As long as you understand the whole system, before you change it using AI, you're good."
Related Reading
- Knowledge Graph - The layer under Blitzy's agents
- Context Engineering - Task context packets as context engineering
- Human-in-the-Loop - Plan approval before agents run