From Volodymyr Bilonenko comes GeoSQL, an AI agent skill for working with geospatial data in PostGIS, BigQuery1, Snowflake2, and Wherobots3. GeoSQL can be used with Claude, Codex, and GitHub Copilot, locally or self-hosted, and in conjunction with Dekart if one wishes to do so.
GeoSQL runs an agent loop, with a map in it, that encompasses these steps:
- Discovery: With GeoSQL, the agent explores the warehouse metadata (tables, columns, types). This works with OMF shares on BigQuery and Snowflake, and with private data in PostGIS, BigQuery, Snowflake, or Wherobots.
- Writing SQL: The agent writes spatial SQL suitable for the database engine in use.
- Cost check for BigQuery: BigQuery queries are dry-run to estimate data scanning costs an designed tighter, if necessary.
- Validating geometries: Using GeoSQL, the agent computes total area (for polygon data) or total length (for linear data) as a sanity check and cross-checks these numbers against domain knowledge.
- Analysing the map: When available, the agent renders the result through Dekart, looks at the rendered image, and corrects mistakes a text-only agentic loop might overlook.
Volodymyr credits the map-in-the-loop for an increase in quality over similar approaches:
Dekart [with GeoSQL] gives the agent a map harness, so [the AI] loops on SQL → Map, checking every result against real geography until it holds up, and fixing itself when it doesn’t.
GeoSQL is open-source under the MIT license. Discussion on Hacker News can be found here.
Footnotes
Google’s data warehouse (DWH) Platform-as-a-Service offering.↩︎
A cloud-based data warehouse (DWH) Platform-as-a-Service offering.↩︎
Wherobots is a cloud platform for large-scale geospatial data processing and analytics, built on Apache Sedona.↩︎