GeoRetina
All posts

Features6 min read

Introducing Cloud AI GIS Agents in GeoRetina AI 2.0

AI agents that take on custom GIS analyses and long-running workflows: they plan the work, write and run the code on cloud machines, and have their results checked before they reach you.

Shahab JozdaniFounder, GeoRetina

Ready-made analyses answer common questions. Real projects rarely stop there. There's always a step that doesn't fit a template: combine three layers in a particular way, measure exposure by your own zones, rank candidate sites by rules your team agreed on, clean and restructure a dataset before anyone can use it. That work has usually meant an analyst, a set of scripts and a long afternoon in a desktop GIS.

GeoRetina AI 2.0 adds Cloud AI GIS Agents for that work. You describe what you need in chat. An AI agent plans the analysis, writes the code, runs it on cloud machines, checks what came out and corrects course as it goes. GeoRetina then verifies the results before they reach you. In the app, this runs as a Dynamic GIS analysis.

Your data

  • Your filesGPKG · SHP · TIFF
  • Earlier resultsFrom past analyses
  • Study areaDrawn or selected
Cloud agentDone
  1. 01Plan
  2. 02Write code
  3. 03Run
  4. 04Check
  5. 05Deliver

workflow.py

zones = gpd.read_file(path)zones = zones.to_crs(utm)zones.geometry = zones.buffer(200)hits = gpd.sjoin(bldgs, zones)

Results

  • Map layersVector results
  • FiguresCharts and maps
  • FilesGPKG · CSV · TIFF
The agent plans, writes and runs code in an isolated cloud workspace. GeoRetina checks what it delivers before it reaches you.

Built for custom GIS workflows

The agent works the way a GIS analyst with a scripting environment does, using the open-source Python geospatial stack. That makes it a general tool for custom work rather than a fixed menu of analyses:

  • Custom spatial analysis: overlays, proximity and buffer analysis, spatial joins, zonal statistics, and ranking and site selection, combined into the multi-step workflow your question needs.
  • Data preparation: clean, reproject, merge, restructure and convert data, or edit attributes in your own datasets and get them back ready to use.
  • Derived products: analysis grids, indicators, summary tables, charts, maps and written reports.
  • Building on GeoRetina's analyses: take the results of a land cover, flood or other GeoRetina analysis further in the same request, without exporting anything.
Try asking

  • Using my uploaded parcels and roads and the flood risk layer from our last analysis, find parcels over 2 ha that sit outside the 100-year floodplain and within 500 m of a road. Rank them by size and return a GeoPackage, a summary table and a map of the top 20.

    Multi-criteria site selection

  • Map land cover for this watershed in 2019 and 2024, build a class-by-class change matrix, work out how much forest turned into each other class within each sub-catchment in my boundaries file, and write a short report with charts.

    Chained analyses, change detection and reporting

  • Overlay the flood depth raster from our flood analysis with my building footprints and road network. For each neighbourhood in my boundaries file, count the buildings and kilometres of road under more than 0.5 m of water, then return a ranked table, a choropleth map and a CSV.

    Raster and vector exposure by zone

  • Build a 250 m hexagon grid over the study area. For each cell, summarise mean NDVI from the vegetation analysis and the share of built-up land from the land cover map, flag the cells with the least vegetation and the most built-up land, and export them with a map.

    Grid-based indicators from several results

  • My uploaded geodatabase has parcels in three coordinate systems with inconsistent zoning codes. Reproject everything to the local UTM zone, standardise the codes with the lookup table in my CSV, repair invalid geometries, and return a clean geodatabase with a log of every change.

    Data cleaning and standardisation

Long-running analyses, handled in the cloud

Heavier analyses take longer than anyone wants to keep a laptop open for. The agent runs on cloud machines, not in your browser, so long jobs keep going in the background while you get on with other work.

A single run can take many steps. The agent writes a program, runs it, inspects the output and moves on to the next step, saving its workspace along the way. Preview images let you follow its progress, and when it finishes, the results are waiting in your chat. If a run reaches its time limit, you keep the partial results it produced rather than starting over.

Your data in, results out

Start from the files you upload in chat, the results of earlier GeoRetina analyses, or a study area you draw or select on the map. The agent reads the common vector, raster and tabular GIS formats.

What comes back depends on the work: layers you can explore on the map in GeoRetina AI, figures and tables, a written summary of the method and findings, or edited datasets in standard GIS formats your team can open in its own tools.

How a run works

  1. Plan. GeoRetina AI turns your request into a plan: the objective, the results you expect back and any constraints. The inputs are fixed when the plan is made, so the run works on exactly the data in the plan.
  2. Write and run code. The agent writes Python and runs it on a cloud machine, one program at a time. After each program it reads the output, inspects the files and corrects its own mistakes before moving on.
  3. Check. When the agent says it's done, GeoRetina runs its own verification program against the delivered results, in a fresh workspace. That program is written by our team, not by the agent. If a check fails and the problem can be fixed, the agent gets a chance to fix it.
  4. Deliver. You get the results to explore and download, with a written summary of the method and the findings.

Checked before it reaches you

Code written by an AI can run without errors and still be wrong, so the agent doesn't grade its own work. GeoRetina's checks confirm that every delivered file opens and holds data, that its coordinate system and coordinates are sound, that geometries are valid, and that results actually cover the study area.

Counts, sums, areas and lengths the agent reports are recomputed from the delivered files. A figure that matches is shown as a metric. Any other figure is labelled as reported by the scripts and not independently verified, so you always know which numbers have been checked. The results also come with a downloadable verification receipt that records the checks.

What the checks cover

The checks make sure the results are sound and the arithmetic adds up. They don't decide whether a method is right for your question. That judgement stays with you, and the written summary of the method is there to help you make it.

Secure by design

Each program runs in its own isolated container on a cloud machine, with no network access and no credentials. Your inputs are mounted read-only, and the agent can't download data or call outside services: it works only on the data attached to the run.

Where you can use it

Cloud AI GIS Agents run in Super mode in the GeoRetina AI web app, where they can work on your uploaded files and earlier results. You can also start a Dynamic GIS analysis through API v1 or the MCP connector in Claude, ChatGPT and other MCP apps, where it works from a study area or from analyses in the same request. Runs use credits for the data, model time and compute they take.

If you don't see Dynamic GIS in Super mode yet, contact us and we'll enable it for your account.

Also new in GeoRetina AI 2.0

Cloud AI GIS Agents arrive with two other features:

  • GeoAI Memory: agents remember your preferences, project decisions and past analyses, so your next task builds on earlier work. Turn it on under Settings → Memory.
  • Scheduled Tasks: set an analysis to run daily, weekly or monthly in the background, with an email when each run's results are ready. Available on paid plans.

Watch the GeoRetina AI 2.0 demo (2:40) to see all three at work.

More from the blog