Where a human approval belongs in an ML pipeline, and where it doesn't
Put a person in front of the leg that spends money or changes production, and show them the evidence. Let data decide everything a number can decide. How Cairn's gates are built around that rule, and what we deliberately keep away from approvers.
Read the post →Fine-tune with a cost gate
How Cairn's fine-tune pipeline prices a run before a GPU starts, checks the trained model against held-out data, and waits for a person before anything is promoted. The commands to run it are at the end.
Read the post →From a DAG of scripts to a governed pipeline
If your ML work already runs as a DAG of scripts, Cairn doesn't replace the tools that run it. It adds what a DAG doesn't record or enforce: cost caps, data and human gates, one record per leg, and a trace you can replay.
Read the post →
Run it on your own infrastructure.
Apache 2.0 licensed, self-hosted. The quickstart gets a control plane with demo data running; a call gets your first pipeline onto it.