cairn is an open-source, self-hosted control plane for governed ModelOps. Fine-tune, evaluate, promote and serve models as commands that are cost-capped, policy-gated and audited, with a person approving the spend. Apache-2.0.
Fine-tuning a model is easy to start and hard to run responsibly: a script on a GPU box, nobody approving the spend, no record of which data went in, and “is it better than the base model?” left to a guess. cairn turns each fine-tune into a governed pipeline: data pinned and checked, test data kept out of training, a person approving before GPU time is spent, cost estimated and capped, the box started and stopped for you, every step recorded for replay and audit. Open source (Apache-2.0) and self-hosted. Here we share what we learn running real fine-tunes this way, with real numbers.