This isn't a talk about AI-assisted coding. It's about what happens when LLM responses drive your application's behaviour; when the output gets parsed, validated, and executed by your business logic.

While building BraidFlow, I discovered that the industry's documented problems aren't solved by better models or bigger context windows. Context drift in multi-turn conversations isn't just a prompt engineering challenge, it's an architectural one. Structured output reliability doesn't need JSON mode; schema field ordering and validation-driven retry logic function as instruction. And cost optimisation isn't about getting the one-shot right with the best performing models; informed retry prompts with cheap models can work just as well.

I'll share the benchmarking data that surprised me, the architectural patterns I tried along the way, and the prompt engineering insights that actually worked. You'll see real code, real failures, and the decisions that finally worked.

No theory. No hand-waving. Just lessons from shipping features that had to work.