AI shouldn’t be one big expensive model doing everything. Build a two-layer stack: a cheap, open-weight “engine” for execution, with frontier models as “steering” for judgment and edge cases. Own your costs, avoid lock-in, and let quality go up as prices go down.