Session

Scaling Geospatial Analytics at S&P Global Energy: From Billions of Points to AI-Powered Map Agents with Databricks

Overview

Experience	In Person
Track	Data Warehousing
Industry	Energy & Utilities, Enterprise Technology, Financial Services
Technologies	Genie
Skill Level	Intermediate
DOWNLOAD SESSION SLIDES

At S&P Global Energy, we transformed our geospatial analytics architecture to handle billions of records with query results in seconds — removing the need for separate geospatial databases and reducing costs using our Oil & Gas well production data.

Architecture evolution:

Unified single-source-of-truth on Delta Lake without redundant geospatial systems
Databricks Spatial SQL with native spatial types enabling performant queries
Databricks SQL Serverless delivering on-demand H3 aggregations

Interactive Databricks app demo for large-well datasets.

What’s next – AI geospatial agent (work in progress):

Natural language map interactions powered by Agent Bricks
Genie-driven text-to-SQL for dynamic map layer creation
50+ analytical tools like chart generation from spatial queries

Attendees will learn practical patterns for scaling geospatial workloads, leveraging Spatial SQL with native spatial types, H3 indexing, and see how AI agents are reshaping spatial analytics workflows.

Session Speakers

Hubert Boguski

/Software Engineer
S&P Global

Michael Johns

/Lead Geospatial Product Specialist
Databricks

Full Summary

Talking to Spatial Data: How S&P Global Built an AI-Powered Map Agent on Databricks

Geospatial users rarely stop at the handful of questions a dashboard anticipates. S&P Global Energy addressed that gap by pairing billion-point spatial SQL with a natural language map agent so users can ask ad hoc questions of governed well data and see answers rendered directly on the map.

The work builds on prior consolidation of geospatial workloads on the Databricks Platform, removing a separate spatial database and its ETL overhead. With one source of truth and high performance in place, the next step was enabling open-ended, conversational analysis without sacrificing governance or speed.

From curated dashboards to open-ended spatial exploration

S&P Global ingests source data once into its Advanced Data Workbench, indexes it with H3, and queries it in place using spatial SQL. The approach reduces duplication, lowers cost, keeps data fresh, and delivers single-second queries across a billion points. Curated dashboards for North American oil and gas wells ride on that same layer.

Fixed dashboards excel at known questions but struggle to cover every operator comparison, production metric, or geographic slice. The team's goal shifted from pre-answering to enabling users to ask their own questions in natural language while staying within the same governed, high-performance data tier.

How the AI-powered map agent operates

A KeplerGL front end renders H3-aggregated wells and progressively drills to finer resolutions as users zoom. Beside standard interactions, a Genie-powered chat accepts prompts such as "show the H3 centroids" or "build a box plot of cells with the highest well counts." Prompts are translated to spatial SQL, executed against serverless Databricks SQL, and returned as either new map layers or linked visualizations that can cross-filter the map.

Every response is just another governed SQL query, so users can compare operators, spotlight highest-average-production zones, or generate histograms without waiting on new dashboards. The conversational analytics experience preserves conversation history, allowing users to refine incorrect results rather than restart.

Platform capabilities that enable the pattern

Spatial SQL has reached general availability with native geometry and geography types, roughly 100 ST functions, and about 35 H3 functions available across Databricks compute, including Databricks SQL. Spark frameworks such as Sedona and partner toolboxes from Carto and Esri can complement the built-ins. AI/BI Dashboards map support is also GA for point and choropleth visualizations.

The app stitches together a small set of Databricks services. Agent Bricks handle natural language to SQL. Databricks Model Serving, moving to Unity AI Gateway, brokers LLM calls such as suggested follow-ups. Unity Catalog governs data access and the agent experience. The thin server is assembled with AppKit, a TypeScript SDK that wires Genie, analytics, and serving plugins, while the browser uses KeplerGL, DuckDB, and Wasm.

Why governed conversational maps matter to customers

Value shifts from delivering static answers to delivering the ability to ask. By keeping data, spatial functions, governance, and LLM interactions within one boundary, S&P Global gives customers genuine exploratory power over trusted energy data without duplicating pipelines or exposing sensitive content. A community app and companion blog are planned so teams can try the same pattern on their own data.

Turning trusted geospatial data into a conversation

Letting users talk to a governed map changes the pace of insight. With spatial SQL at scale and an agent that translates questions into queries, teams move beyond static dashboards to interactive exploration, while keeping performance and control intact.

FAQ

No. Native geometry and geography types, spatial SQL, and H3 functions run directly on Databricks compute, including Databricks SQL. S&P Global eliminated a separate spatial database, reducing ETL cost and improving data freshness.

Dashboards answer expected questions such as well counts by H3 cell. The Genie-powered map agent lets users pose new questions in natural language, which are translated to spatial SQL and returned as new map layers or charts without engineering work.

Unity Catalog controls data access, and Databricks partnerships with foundation model providers keep conversations within the customer's security boundary. The team is moving LLM calls to Unity AI Gateway for centralized, governed model access.

Agent Bricks perform natural language to SQL translation. Spatial SQL and H3 functions execute in the lakehouse. Model Serving handles LLM prompts, with a planned move to Unity AI Gateway. The front end uses KeplerGL, DuckDB, and Wasm, and the server is assembled with AppKit.