2025 FOSS4G NA | When LLMs Meet GIS - Jason Gilman
In this session from FOSS4G NA 2025, Jason Gilman of Element 84 presents When LLMs Meet GIS: Building Reliable Geospatial AI Systems. Jason explores the practical integration of Large Language Models into geospatial workflows, moving beyond the hype to focus on building software that is reliable, cost-effective, and meet real-world user needs.
The presentation introduces two open-source libraries: E84-GDAL-AI-Common, which provides core utilities for LLM interaction and structured data extraction, and Natural Language Geometry (NLG), a library that translates human-readable spatial descriptions—like "within 10 miles of the coast"—into precise GIS polygons. Jason also discusses the trade-offs between autonomous "agents" and deterministic "workflows," the importance of tracing and evaluation suites, and how Element 84 built their own geocoding database to overcome the relevancy and polygon-retrieval limitations of traditional tools.
Highlights:
🛰️ The Reliability Gap: Why rushing to add AI to geospatial tools can compromise software quality, security, and energy efficiency
🔄 Natural Language Geometry (NLG): How to convert vague spatial phrases into a structured tree of operations and final GIS polygons
🧩 Beyond Point Geocoding: The shift from center-point results to full polygon retrieval for complex areas like the Gulf of Mexico
⚙️ Why Build a New AI Library? Moving away from the complexity of LangChain toward the strong typing and simplicity of E84-GDAL-AI-Common
🌐 Agents vs. Workflows: When to use autonomous AI agents for unpredictable human queries and when to use prescriptive pipelines for efficiency
🏢 Evaluations & Tracing: Using graph edit distance and tools like LangFuse to measure LLM performance and identify token bottlenecks
🏗️ Custom Geocoding Databases: Leveraging Who’s On First and Natural Earth data to provide context-aware results for ambiguous place names
For more content like this check out www.projectgeospatial.com
#Geospatial #FOSS4G #LLM #AI #GIS #NaturalLanguageProcessing #Geocoding #OpenSource #SoftwareEngineering #DataScience #Element84 #MachineLearning #ProjectGeospatial #LangChain #Pydantic
