Choosing an AI stack without overengineering it

Teams often overbuild their first AI stack because they expect scale, multi-model routing, and complex evaluation from day one. In practice, most projects need a simpler baseline first.

I prefer to choose the minimum stack that supports ownership. That usually means one model provider, one place for prompts, one deploy target, and one observable request path.

Questions worth asking early

• Who will maintain the prompts and contracts after launch?
• Where does evaluation actually need to happen?
• Which failure mode is more dangerous: bad output or delivery friction?

If the team cannot explain why a layer exists, it probably should not be there yet.