Fine-tuned models for when off-the-shelf isn't enough.

We start with evals, not model code. Only once we can measure quality improvement do we invest in fine-tuning. This saves significant GPU cost and ensures you get a model that actually performs better in production.