AI agents that build, run, check and fix governed data pipelines, then answer the questions the data was for. You write one sentence. They ask one question back.
Each stage is backed by a skill: reviewable knowledge the agent loads before it acts. Skills are data, not code.
One card per system: auth, limits, watermarks and the few questions only you can answer. Four connectors proven end to end, two scaffolded, 31 more cards ready.
Star schemas, SCD2 history and metrics defined once. You approve the design before any table exists.
Full, incremental, updates, deletes, drift and re-runs, each checked against the lineage log, not task status.
Read-only SQL, a Python kernel and published dashboards. Every number ships with its query and source.
A pipeline is a typed contract with a small set of allowed choices. The guarantees live in the engine, so they hold no matter who wrote the contract.
Invalid options are rejected when the contract is built, before anything runs.
Lineage, drift detection and error capture run for every pipeline. A contract can't switch them off.
plan() shows exactly what will change. Writes, runs and publishing wait for a person to approve.
Done means Airflow, the lineage log and the warehouse row counts all agree.
A new source is three small files. A new warehouse is one adapter. No vendor lock-in.
Airflow, PostgreSQL, Superset and JupyterHub in your own cloud. Credentials never pass through chat.
Every table goes through extract, validate and store, and the engine records each step. "It ran" is never the claim. "It reconciles" is.
validate_sql must return 0| table | extracted | validated | stored | warehouse | |
|---|---|---|---|---|---|
| Run 1: stopped, not certified | |||||
| companies | 1 | 1 | 1 | 1 | ✓ |
| filings | 40 | 40 | — | — | ✕ |
| facts | 6,247 | 6,247 | — | — | ✕ |
| documents | 43 | 43 | — | — | ✕ |
| Run 2: after the fix | |||||
| companies | 1 | 1 | 1 | 1 | ✓ |
| filings | 40 | 40 | 40 | 40 | ✓ |
| facts | 6,247 | 6,247 | 6,247 | 6,247 | ✓ |
| documents | 43 | 43 | 43 | 43 | ✓ |
odf_error_log said why. The agent fixed the contract and re-ran. Run 2 certified only when every count agreed.Governance columns on every row trace any dashboard value back to its run and source.
odf_job_idthe run that wrote itodf_row_keythe business keyodf_row_hashcontent fingerprint for change detectionodf_extracted_atfreshnessodf_deleted_atwhen it left the sourceThe pipeline, the notebook and the dashboard all read the same certified data, with lineage back to the source.
Per-table tasks, plans and approvals before deploy, and agent-led debugging with evidence.
Watch as an engineerMetrics defined once, dashboards with definitions and lineage, and the SQL behind every answer.
Watch as an analystNotebooks on the same workspace and kernel as the agent, so its work can be re-run.
Watch as a data scientistEvery action approved and audited, one governance standard, and answers with their evidence.
Watch as a business leaderBring one system and one question. We'll take it to a certified dashboard in one session.