product system
Atlas
A data change intelligence agent that traces the downstream blast radius of a schema change, ranks the risk deterministically, and holds every destructive action behind a human approval gate.
Sole author — agent architecture, lineage engine, risk ranking and interface
Atlas answers a question data engineers ask before every column drop: what breaks if I do this? It walks a lineage graph from the target column, collects the dependent dbt models, dashboards, ML features and scheduled reports, and reports what each one is worth and who owns it.
The architecture separates judgement from action. A Gemini model interprets the request and calls read-only tools during analysis; write tools are simply not available in that phase. Severity is calculated in Python from tier and ownership signals rather than generated by the model, so the risk badge cannot be hallucinated. Execution happens only after a human approves, and every action is written to a timestamped change log.
The Fivetran tool layer mirrors the official MCP server's tool names and response shapes against an in-memory fixture. That is a deliberate, documented choice: there are no Fivetran credentials behind the demo, and the README says so.
What this record says
Problem: Dropping or renaming a data column can silently break dashboards, models and reports that nobody remembers depend on it.
Audience: Data engineers and analytics teams making schema changes against shared warehouses.
Contribution: The whole system: the two-phase agent loop, the lineage engine, the deterministic severity ranker, the Fivetran-compatible tool layer, the approval gate and change log, and the Streamlit interface.
Status: Public repository with an MIT licence and a live Streamlit deployment.
System decisions
- Calculate severity in code, never in the prompt, so the model cannot invent a risk level
- Withhold write tools entirely during the analysis phase rather than instructing the model not to use them
- Require explicit human approval before any change executes, and log it with a timestamp
- Ship a response cache so the demo works when the free model quota is exhausted
Next improvement: A visual lineage graph and rollback for an executed deprecation.
Evidence and sources
31 commits — the only contributor to the repository Open source
Built for the Google Cloud Rapid Agent Hackathon 2026, Fivetran partner track - The Fivetran tools run against curated in-memory fixtures rather than live credentials. This is stated in the project README. Open source