Khartis: Sciences Po's thematic mapping tool

Sciences Po's Cartography Lab designed a complete rewrite of Khartis, its mapping tool for researchers and journalists, and I developed it. Datasets of up to millions of rows become print-quality maps in the browser, even offline, with no data sent to a server.

Client: Sciences Po
Role: Senior full-stack developer, in charge of all development
Duration: 18 months, May 2025 to October 2026 (one engagement, three projects)
Setting: Freelance, via RECSI Group, fully remote

001

001·Context

Version 3 of the Lab's mapping tool

Since 2016, the Cartography Lab (Atelier de cartographie) has offered Khartis to researchers and journalists who need thematic maps without being cartographers. Version 3 is one of the three projects of my Sciences Po engagement, carried out with the Lab's team.

002

002·What makes it new

A complete mapping tool that runs entirely in the browser

The Lab's design puts every step of map-making on the user's machine, from import and analysis to projection and rendering. My work was to make it run in a web page, with large datasets.

Data analysed where it is opened
DuckDB, compiled to WebAssembly, runs in a Web Worker with its spatial and Parquet extensions. Every step runs in the browser, and no server ever receives the user's data.
- DuckDB-WASM in a Web Worker
- CSV, GeoJSON, Shapefile, GeoParquet
- No data sent to a server
A binary pipeline from the query to the GPU
Query results travel as Apache Arrow and GeoArrow columns straight to Deck.gl, which draws them with WebGL2. Projections from d3 give print-quality maps, a MapLibre mode handles tiles and the globe, and basemaps come from a versioned GeoParquet catalogue.
- Apache Arrow · GeoArrow
- Deck.gl on WebGL2
- MapLibre GL
- National projections
The Lab's cartographic vision, implemented
The Lab designed the tool and its map semiology, Antoine Rio its interface, and Thomas Ansart the Lab's open-source libraries it builds on, such as projection suggestions and palettes. I implemented every map type and classification, from natural breaks to nested means.
- Design: Atelier de cartographie
- Interface: Antoine Rio
- Open-source libraries: Thomas Ansart
- Development: Jean-Baptiste Thery
Installable, offline and open source
Khartis installs as a progressive web app and keeps working offline once its basemaps are loaded. Projects are saved in the browser and exported as SVG, JPG up to 4K or GeoPackage. The app, in French and English, is open source under the MIT licence.
- PWA, works offline
- SVG and 4K JPG export
- MIT licence
- French and English

003

003·Approach

How I worked on Khartis

Brief and my role: The Lab supplied the design, the specification, the semiology and the cartographic algorithms. I owned the development, data processing included.
Constraints: Datasets of up to millions of rows, processed without a server, in an app that works offline.
Technical decisions: I implemented the Lab's architecture: DuckDB-WASM and Apache Arrow for the data, Deck.gl and MapLibre for WebGL maps, in a SvelteKit and Svelte 5 PWA on Carbon components.
Checks and releases: I set up automated tests, documentation and a CI/CD pipeline with GitHub Actions.

004

004·Deliverables

What I delivered on Khartis

In-browser data processingUp to millions of rows analysed on the user's machine with DuckDB-WASM
WebGL map renderingThematic maps drawn with Deck.gl and MapLibre, implementing the Lab's cartographic algorithms
Offline applicationA progressive web app that keeps working without a connection
Architecture and deliverySvelteKit, Svelte 5 and TypeScript, with tests, documentation and CI/CD deployment

005

005·Technologies and screenshots