Agentic Point Clouds
What does putting scanned data behind an API actually mean? With LumiDB, two short prompts took an AI agent from a 3D point cloud to an overview of Helsinki’s Hotel Clarion and a measurement between a nearby crane and the hotel: about 55 metres, confirmed manually in the LumiDB Viewer. The user didn’t need to locate the data, extract features, or learn specialist tools. Here’s how LumiDB’s API and MCP integration make this possible and what it could mean for getting useful answers out of huge geospatial datasets.
Written by
Published on
Do not index
We talk a lot about how LumiDB puts scanned data behind an API, blah blah. But what exactly does this mean in 2026? Well, here’s something we’re working on.
The first image is the result of a single prompt, word for word: “Give me a nice overview shot of Hotel Clarion, Helsinki, in PNG format.” I get this nice SimCity-esque (*) pic, which is kinda cool, I guess, but maybe not all that useful.
OK, so a follow-up prompt: “What is the distance between the tip of the crane and the nearest hotel tower?” And, after a bit of processing, the agent spat out: “About 55 m from the long boom tip to the nearest hotel tower façade—a 3D straight-line estimate from the September 2022 point cloud.” (With the em dash and all!) I then fired up the LumiDB Viewer and measured it manually to confirm, and that indeed is the distance. The real-world application possibilities of such an analysis are obvious. (Think encroachment analysis!)
.png?table=block&id=3f411b43-120a-80a6-9d56-ccac14ea3e34&cache=v2)
Now, it’s easy to brush this off as just something AI “just does these days.” But the input data was just point a bunch of 3D points with some other attributes. LumiDB parsed a ton of metadata from it and exposed it via MCP to the AI agent. The agent was then able to a) locate the point data corresponding to the specific hotel building, b) pull out 2D visualizations with an API call for feature analysis, c) identify the exact spatial location of the tip of the crane near the hotel, and d) compute the distance between the two. Two short prompts and a couple of minutes. That's it.
The dummy user didn’t have to know anything about how to extract features, analyze distances, or even use any tools. Just vibe some prompts and get an actually (well, more like hypothetically) useful answer back. It’s easy to get jaded by all the AI hype, but I still sometimes find it remarkable what it’s able to pull off.
We’re making LumiDB the system where the value of these huge and expensive geospatial datasets can best be utilized by both humans and AI alike. Check it out, if you’d like to see it in action!
Written by
