
Among developers who build with AI, 67% use it to generate code in areas they already know, but only 20% use it to deploy, operate or troubleshoot production systems, according to the 2026 Developer Survey from Stack Overflow, the widely used programming forum.
That gap between drafting code and trusting it in production was at the heart of what Rakesh Awasthi had to say about modernization.
Rakesh, a principal software engineer at Mastercard who works on large-scale data platforms, joined me for his second “In the Arena” conversation at AI Infra Summit, held Sept. 15-17, 2026, at the Santa Clara Convention Center in California. He had just come off the stage after his session, “Generative AI as the Engineer: A Practitioner’s Guide to Full-Lifecycle System Modernization.”
The Workload Moved to Validation
Rakesh’s case for why modernization deserves attention right now starts with where the work happens. Teams still follow the standard software development life cycle, he said, but the effort inside it has moved.
“Generative AI has shifted the major workload from actual development to validation,” he said.
That matters to the people building AI infrastructure. Infrastructure engineers need to understand how data engineers will use what they set up, and data engineers need to know where that infrastructure can support them best.
“It’s not only about LLMs and token optimization; it’s actually more of a validation problem,” he said.
Operators modernize for many different reasons, not only because legacy systems stop operating well.
“You are not always modernizing because the old legacy system is not working,” Rakesh said. When the landscape shifts and new governance or security principles arrive, the modern system has to absorb them while still matching the old system’s results.
The biggest initial investment goes into understanding the old system, Rakesh said. A legacy system that has run for 5, 10 or 15 years may carry undocumented code changes, unwritten requirements and open bugs, while the people who could explain them may have moved on or retired.
Generative AI can help teams work through that system and question it before moving ahead. What they learn becomes the specification and baseline for proving out the new system. In his view, modernization has become “more of an understanding problem now than it ever used to be.”
Earning Trust Before Production
“Trust comes with the confidence when you have repeatedly proven that the system works,” he said.
Rakesh runs legacy and modern systems in parallel and checks that both produce similar results in data output, performance, security, governance and observability. Sometimes this means intentionally trying to break the system to make sure that the LLM isn’t hallucinating or giving false positives.
“When we have sufficient understanding of that and data points for that and humans have actually validated that result, then I think we’ll be ready to put that in production,” he said.
Business leaders add context about the data, Rakesh added, and an independent perspective that can catch what engineers overlook after seeing the same system for so long.
I asked whether his peers were seeing more inference applications.
“There is definitely an uptick there,” he said.
Where teams once had to decide which modernization projects to prioritize, faster AI-driven work now gives them room to take on a much larger number of them.
That is the exciting part and also a scary one, because teams may be producing issues at the same speed, he added. He pointed to vulnerabilities and hacks over the past year in code written by AI. Researchers are now counting them. Infosecurity Magazine reported that Georgia Tech’s Vibe Security Radar project tracked at least 35 new common vulnerabilities and exposures (CVE) entries disclosed in March 2026 that were a direct result of AI-generated code, up from six in January.
One use case presented at the summit stayed with him. SewerAI, a Walnut Creek, California, company that applies AI to sewer inspection, uses AI to inspect sewer pipelines and find where the next break may happen. Inspecting the pipes across Santa Clara and San Jose by hand could take more than a decade, and the company has cut years of that work to weeks using consumer-grade cameras.
“That’s not a very hygienic place where a human wants to go and look into everything,” he said.
A year out, Rakesh expects AI applications to reach further into health care, policy and public safety. He also sees validation moving into the models themselves. Six months earlier, even the best models didn’t check their own work, and today models such as Claude Opus 5 do. According to Rakesh, infrastructure teams are working to build security in at that level. “I think some of it will actually start showing up next year itself,” he said.
Rakesh’s focus on validation has direct implications for infrastructure teams. When AI can generate code quickly, the harder work becomes proving that the code is right. Running old and new systems side by side, trying to break the new one and having people review the results all take compute, storage and data pipeline capacity.
His advice to understand legacy code first is the most practical. Using generative AI to surface undocumented changes and open bugs in an old system gives a team a dependable baseline to test the new one against.
Developers already feel the gap between generating code and trusting it in production. The coming year will show whether self-checking models and security built into infrastructure start to close it. To learn more, watch the full podcast.