Ever wondered how the people building the tools we use every day are leveraging AI in their own work? Well, Stephen Toups and David Fowler from Microsoft’s .NET team recently pulled back the curtain on exactly that. In a candid presentation, these developers shared real examples of how they’re using GitHub Copilot, ChatGPT, and other AI tools – not for flashy demos or marketing material, but for actual day-to-day development work on .NET itself.
What’s refreshing about their approach is the honesty. They’re not claiming AI will replace developers or solve every problem. Instead, they show how AI has become a practical tool in their toolkit for everything from brainstorming solutions to automating boring documentation tasks. From solving decade-old bugs to optimizing performance and generating thousands of lines of documentation, their examples offer a realistic look at where AI tooling actually adds value in professional software development.
If you’ve been curious about how to practically integrate AI into your development workflow beyond basic code completion, this is worth your time.
AI for Ideation and Problem Solving
The biggest highlight was David Fowler using AI to tackle a decade-old Kestrel memory issue that was causing out-of-memory errors. Instead of looking for a quick fix, he used ChatGPT to rapidly brainstorm different approaches – essentially treating it like a mini garbage collector problem. The AI generated four potential solutions in seconds, helping him iterate through ideas much faster than traditional methods. This eventually led to a production-ready implementation that’s now going into the main branch.
Enhancing Performance with AI Suggestions
Stephen Toups shared some impressive wins using AI to optimize the .NET regex engine. After providing specific examples and concrete prompts, Copilot suggested optimizing lookaheads – something that had been previously overlooked. This resulted in a 10x performance improvement for regex patterns used in SQL statement parsing. He also used AI to implement LINQ’s new shuffle method, where Copilot suggested reservoir sampling for better performance.
Documentation and Maintenance Made Easy
One of the most practical applications was using AI to tackle the boring but necessary stuff. Stephen used Copilot to generate XML comments for 3,000 lines of cryptic regex interpreter code from 2003. For the Model Context Protocol library, AI generated 6,000 lines of documentation comments, which Stephen then edited down to 4,000 lines – saving days of work and solving the dreaded blank page problem.
Debugging and Issue Resolution
The debugging examples were pretty compelling. Stephen pasted an email about a memory leak in System.IO.Pipes into Copilot, and it identified the exact problem and location in about 45 seconds. While the AI’s proposed solutions weren’t perfect, it pointed him in the right direction for a quick fix. They also showed Copilot Agent creating a complete pull request to fix a networking team issue in just a few minutes.
Live Coding Demo
David demonstrated Copilot in Agent Mode live on stage, adding properties to multiple Azure resources in the Aspire project. The AI scanned the codebase, identified what needed to be done, and made changes to six different files including adding required comments. It even generated the pull request description and reviewed its own work.
Key Takeaways
Both developers emphasized that AI works best when you have some context about the problem area and use it for bouncing ideas rather than just code completion. They see AI as turning developers into “10x developers” by automating the trivial tasks they know how to do but don’t want to do, freeing up time for more creative problem-solving.
Check out the video: