Assess
The scan runs as a dataflow job; the agent narrates duplicates, contamination and PII findings.
Agentic · ships today
Check and repair
a dataset
An agentic check narrates the verdict; a bounded loop repairs and re-checks; every destructive write waits for approval.
Ships as dataset_readiness/readiness
A fused single-pass scan runs on your engine, an agent turns the findings into a verdict and a plain report, and a repair loop deduplicates, decontaminates and redacts, re-checking after each pass. Nothing is written until a person approves.
The pipeline
Approve repairs
Every repair is a destructive write, so the run pauses for a person first.
Repair
Dedup, decontamination and redaction run as one job; a million rows finish in minutes.
Re-assess
The check runs again. The loop is bounded; exhaustion asks a person.
What the record keeps
The scan findings, the narrated verdict, each approved repair with a report of exactly what changed, and the final verdict.
Nothing silent
Each repair reports exactly what it changed.
Heavy work off the control plane
The scan and repair run on your engine, not the control worker.
Deterministic repairs
The fix steps use no LLM and no GPU; the agent only narrates and decides what to propose.
Ships as dataset_readiness/readiness, check, fix and fix-batch. Engine support depends on the deployment.
See how failure can trigger a reviewed improvement cycle →Run it on your own infrastructure.
Apache 2.0 licensed, self-hosted. cairn pipeline run dataset_readiness/readiness