Thoughts

I keep coming back to something I noticed when I was teaching at Forge: the best way to really understand something is to explain it to someone who’s never seen it before. Every team I’ve educated on sovereign cloud patterns has taught me something about my own assumptions.

The gap between “this model supports function calling” and “this model reliably uses function calling in production” is enormous. Spent the week investigating tool-based RAG for Edge-RAG and the smaller models (Phi-3.5, Mistral) still struggle with knowing when to call a tool vs just answering from context.

One thing I’ve learned building for air-gapped clouds: never assume infrastructure parity. The feature that works perfectly in public Azure might silently fail when you can’t reach an external endpoint. Now I start every feature by asking “what does this depend on, and is it deployed everywhere?”