▶ Watch ↗19:43
Conference affiliation: Blitzy
Neeraj Deshmukh is field CTO at Blitzy, where he works with enterprises applying autonomous software development to large, interconnected systems. He previously led the company’s research and engineering in code understanding, agent orchestration and evaluation. His work addresses a problem that code generation alone cannot solve: determining how a change affects the rest of a software system, then testing whether it preserves the required behavior.
Neeraj’s early research addressed the search problem inside speech recognition: finding a likely sequence of words among many acoustically plausible alternatives without exhausting available computing resources. At Mississippi State University, he co-authored research on hierarchical speech decoding with Aravind Ganapathiraju and Joseph Picone. Their 1999 work described a public-domain decoder that combined acoustic, pronunciation and language models, using dynamic programming and pruning to reduce computation while retaining recognition accuracy. The contribution included both recognition methods and software other researchers could use.
He subsequently held technical leadership roles at Dragon Systems and Nuance, developing speech recognition for enterprise and healthcare applications. His speech work also included speaker adaptation: he and Puming Zhan are named inventors on a Nuance patent granted in 2013 for transforming speech features using speaker-dependent adaptation before recognition. Before Blitzy, he worked on Microsoft’s Dynamics 365 Contact Center, bringing advanced speech recognition and agentic AI into customer-service systems.
At Blitzy, Neeraj led research and engineering before becoming field CTO. He co-authored the team’s software-engineering platform paper, which describes two deliberate tradeoffs: accepting longer response times to assemble useful context across a large codebase, and spending more inference compute to validate the resulting changes. His work connects the knowledge layer agents use to understand software with the orchestration and evaluation needed to carry out changes across it.
In his joint AI Engineer presentation with Stephen Chin, Neeraj used Amazon’s public AWS CardDemo application to explain why translating a function is only part of a migration. A COBOL interest calculation sits within a batch process, uses record layouts defined in copybooks, and feeds account updates and transaction records. A Java translation can compile and still pay a different amount if it changes field precision or rounds at a different point.
The example makes several parts of his approach concrete:
In his account of reviewing AI-generated software, Neeraj argues that reviewers need the purpose and validation of a change alongside its diff. Blitzy captures that information while agents work, groups consequential changes for inspection, and delivers a project guide explaining what changed and how it was tested. Human review begins with the plan and continues through the evidence accompanying the resulting code.
This places specific demands on the engineers directing the work. They must define the intended outcome, constraints and acceptance criteria, and provide access to the environment needed to test the result. Neeraj identifies ambiguous requirements and inadequate test access as practical causes of disappointing outcomes. More autonomous implementation increases the importance of specifying what success means and checking the delivered system against it.
▶ Watch ↗19:43
Loading more talks and highlights…