BONUS: How AI Is Reshaping Software Teams From the Inside — Lessons From Google, Meta, and Snowflake In this episode, Dwarak Rajagopal — VP of AI Engineering and Research at Snowflake — shares what he's seeing firsthand as AI agents become part of the software development process. From compressed sprint cycles to automated standups across time zones, Dwarak draws on two decades of building AI infrastructure at Google, Meta, Uber, and Apple to show what's actually changing inside engineering organizations today. From Compiler Engineer to AI Leader — The Thread That Connects Two Decades "In AI, the hardest part isn't just the models itself, it's making them work in real environments where data is messy, fragmented, and governed." Dwarak started his career as an open-source GCC compiler engineer over two decades ago, optimizing hardware performance. He moved into graphics at Apple, then pivoted to AI when AlexNet started running on GPUs around 2011-2012. From there, he built autonomous driving software at Uber, led Meta's PyTorch core framework team bridging research and production, and at Google led AI Frameworks including getting Gemini training on TPUs. The common thread: always working at the intersection of research and production, making powerful technology work in the real world. That focus on real-world application is what drew him to Snowflake — where enterprise data meets AI at scale. AI Is Changing What Engineers Actually Do All Day "Engineers are spending more time on system design, validation, production reliability — and less time doing the implementation itself, because AI is helping that." The shift Dwarak sees is concrete: AI is accelerating development, but the real value comes when it's grounded in enterprise data and context. At Snowflake, teams use tools like Cortex Code, Snowflake Intelligence, and other LLMs to generate code and tests faster — because the friction cost of development has dropped dramatically. Customer example: Whoop, the fitness band company, used Cortex Code with conversational data assistance and agents to reduce development cycles from weeks to hours, freeing teams to focus on high-value work. The End of "This or That" — Try Both, Kill Fast "There's a lot more choices now. You don't have to think about this versus that. Do both and then figure out what is the best." One of the most practical shifts Dwarak describes: teams no longer need to commit to one architectural approach upfront. Because AI reduces the cost of building, teams can pursue two designs in parallel and evaluate both. A concrete example: instead of choosing a cross-platform framework like Flutter or React Native for a mobile app, Snowflake's teams now build native iOS and Android apps simultaneously — one human-led, the other agent-built — at roughly the same speed. But this creates a new challenge: teams have to learn to kill projects faster. When you can build more, you also discard more — and engineers need to detach from "their baby." Smaller Teams, Bigger Output — The Cross-Functional Shift "You could build multiple products now faster with different smaller teams. One back-end person, one front-end person — build vertically end-to-end." Dwarak's teams moved from functional structures (separate backend, frontend, and feature teams) to project-based teams that own the full vertical stack. This isn't theoretical — Snowflake Intelligence was built this way. The result: fewer dependencies, faster delivery, more products in parallel. The tradeoff is coordination cost — more things running in parallel means more decisions to synchronize. Recruiting Has Fundamentally Changed — Systems Thinking Over Syntax "We used to ask an engineer to code a specific search algorithm. Now we ask them to build a whole search system within an hour." Dwarak is clear: fundamentals matter more than ever. Systems thinking, judgment, the ability to work with complex data and production systems — these are what hiring evaluates now. AI handles execution; humans need to define problems clearly and ensure systems behave at scale. For junior engineers, the news is encouraging: onboarding is faster because team-specific skills are codified and shared, and the barrier to building end-to-end systems has dropped. "Learning by building is more true than ever now." Monday Planning, Friday Demos — The Compressed Sprint "You basically decide what to do on Monday, and you're testing together as a team on Friday and getting the feedback for the next week." Daily work has transformed at Snowflake. The traditional multi-week sprint has compressed to a single week: Monday planning, Friday team demos and testing. Standups still happen — but faster, sometimes multiple times per day. For distributed teams across Bay Area, Seattle, and Poland, an automated skill scans each day's code changes and posts a summary in a shared Slack channel — so the next timezone knows exactly what happened without waiting for a meeting. This solves one of the oldest problems in distributed development. The Road to Lights-Out Codebases — Governance, Observability, Reversibility "Can agents take actions? Which of these actions cannot be taken back? You need the concept of committing actions or rolling back." Building on the "lights-out codebases" concept from Philip Su's episode, Dwarak agrees the direction is clear — agents are already writing more code than humans in some contexts. But enterprise adoption requires governance, observability, traceability, and reversibility of agent actions. The shift from "AI as a tool" to "AI as part of the system" is happening now, with the focus moving from getting answers to enabling actions at scale. What Most People Get Wrong About AI in Software "It's very easy to build prototypes, even end-to-end systems. But it's very hard to get it working in enterprises where the data is so messy." The gap between demo and production is where most organizations hit the wall. Enterprise data is scattered across invoices, factory outputs, and dozens of systems — combining it meaningfully for AI to generate insights and actions is the real challenge. This is different from the "AI will replace developers" narrative. The bottleneck isn't code generation; it's data integration, governance, and controlled execution at scale. About Dwarak Rajagopal Dwarak Rajagopal is VP of AI Engineering at Snowflake, where he leads the Cortex AI and AI Research teams. Before Snowflake, he led Google's AI Frameworks and On-Device ML teams (including Gemini), ran Meta's PyTorch Core Frameworks team, and built autonomous driving software at Uber. Two decades of shipping AI at the companies that define the field. You can link with Dwarak Rajagopal on LinkedIn.
