MLOps.community Getting AI Apps Past the Demo // Vaibhav Gupta // #319
51 snips
May 30, 2025 Vaibhav Gupta, CEO of BoundaryML and BAML creator, shares insights from his decade in AI performance optimization at giants like Google and Microsoft. He critiques current prompt engineering, advocating for organized coding practices to enhance AI reliability. The discussion spans the evolution of web development and AI integration challenges, emphasizing the need for programming languages that support large models. Gupta also introduces BAML, a language designed for seamless integration, showcasing its promising applications in sectors like government and healthcare.
AI Snips
Chapters
Transcript
Episode notes
AI Code Feels Like Demo Code
- AI pipeline code often resembles quick throwaway demo code rather than production quality code.
- Lack of rigor and maintainability in current AI codebases hinders the transition from demo to production.
Prompt as Verifiable Code
- Current prompt engineering uses unstructured English strings, making it hard to verify and debug.
- Treating prompts like code with syntax highlighting and live preview greatly improves iteration speed and reliability.
Verifiable English Explained
- Verifiable English means prompts act like functions with inputs and outputs that can be tested.
- Treating LLMs as calculators with operators (prompts) allows easier debugging and prompt iteration.