Code generation is a leading use case for generative AI, but much of the focus remains on building entire applications from prompts — often at the expense of reliability, interoperability, and standards compliance.
Given the well-known tendency of large language models to hallucinate, AI-generated code must be validated and deterministic to be useful in real-world systems. Instead of treating LLMs as full-stack developers, we should leverage them as assistants for creating standards-compliant tools — tools that generate predictable, testable outputs that slot into existing ecosystems.

This talk explores a spec-first approach to AI-assisted backend development. I’ll share how focusing on service design, not end-to-end app generation, enables better integration, repeatability, and collaboration. I’ll conclude with a demo of Foundry, a tool that consumes OpenAPI specifications to generate production-ready backend services, illustrating how AI can support — not replace — rigorous system design.