AI / Geospatial / Professional internship
EpiWise Geospatial LLM Chatbot
LangChain/LangGraph agent for interacting with epidemic-risk geospatial maps.
- Role
- AI intern / agent developer
- When
- Mar 2025 — Aug 2025
- Status
- professional project
EpiWise is a geospatial epidemic risk platform providing risk projections, datasets, and an integrated decision platform. During the Geomatys internship, the AI chatbot work focused on making complex geospatial models easier to explore through natural language. The agent could operate application controls, reason over geospatial API outputs, capture map views, and report results in a user-friendly way.
The platform
EpiWise turns satellite and epidemiological data into risk projections on an interactive map — this is the application the agent operates.
Talk to the map
EpiWise's models are powerful but dense — the agent I built at Geomatys lets non-scientists drive them in plain language. A typical session:
Show me the dengue risk projection for southern France in August.
Switching the map to the dengue risk layer and setting the date to August. The southern coastal departments show elevated risk — here is the current projection.
changed layer → dengue-risk · set date → 2025-08
Why is the risk higher around Montpellier?
I queried the WFS layer behind this projection: the model weights summer temperature, standing-water indices and historical vector presence. Around Montpellier all three indicators are elevated.
tool call → WFS feature query
Great — send me a report of what we just looked at.
Done. The report includes our discussion history, screenshots of each map state and the model parameters used.
generated report → screenshots + discussion summary
Under the hood
- ReAct agent — LangChain/LangGraph loop reasons about each request before acting.
- Tool calling — map layers, dates and model parameters are real application controls the agent operates.
- WFS / WMS — geospatial standards give the agent ground-truth context for its answers.
- Reporting — screenshots + discussion history compiled into a final document.
- MCP / A2A — explored as the long-term protocol layer for maintainability.
Features
- Natural-language map interaction
- Layer/date/model parameter control
- Screenshot capture
- Discussion-history based reporting
- Geospatial API tool calling
- Non-expert user assistance
What I did
- Designed a ReAct-style agent with LangChain and LangGraph.
- Connected the agent to geospatial app controls and APIs.
- Added tool workflows for map layer/date/parameter changes.
- Integrated WFS/WMS tool usage for better response context.
- Added screenshot and final report generation flows.
- Explored MCP/A2A protocol directions for future maintainability.
Project timeline
Professional AI internship project: LangGraph/LangChain geospatial agent integrated into EpiWise map workflows.
-
Mar 2025 Planning / Research
Internship and chatbot project started
Started the Geomatys AI internship and the EpiWise geospatial chatbot project.
-
Apr 2025 ~ Development
ReAct agent implemented
Implemented a LangChain/LangGraph ReAct agent capable of interacting with geospatial APIs and application state.
-
May 2025 ~ Development
WFS/WMS tool integration
Built LLM tools to interact with geospatial WFS/WMS APIs and gather context for better map-based answers.
-
Jun 2025 ~ Development
Map actions and reporting workflow
Enabled the chatbot to change layers, dates, model parameters, take screenshots, and report results to non-scientific users.
-
Jul 2025 ~ Planning / Research
LLM protocol exploration
Started implementing future-oriented LLM protocols such as MCP and A2A for maintainability and extensibility.
-
Aug 2025 Release / Delivery
Internship delivery and handoff
Delivered the professional internship work and contributed to the EpiWise project context.
Built with
- Python
- LangChain
- LangGraph
- Mistral
- ReAct
- WFS
- WMS
- Geospatial APIs
- MCP
- A2A