Talk, research and experiments for "Making deck.gl AI-Ready: Declarative Specs as an Agent-Native Interface for Maps", Open Visualization Collaborator Summit, ETH Zürich, 9 September 2026. Speaker: Javier de la Torre (CARTO).
- Slides: https://jatorre.github.io/deckgl-ai-ready/ (reveal.js, press
Sfor speaker notes) - Outline with speaker notes:
talk/OUTLINE.md
| Path | What |
|---|---|
research/experiments/RESULTS.md |
Experiment 1: what 10 models recommend for maps, how current their deck.gl knowledge is, whether the MapLibre relationship confuses them, and whether they can write @deck.gl/json. Raw answers in results/. |
research/experiments/mapbench/ |
Experiment 2: 7 frontier models (Opus 4.8, Fable 5.1, GPT-5.6 Sol, GPT-6 Astra, Gemini 3.1 Pro, Gemini 3.8 Flash, GLM 5.3) build 4 specified single-page deck.gl maps plus one open brief; rendered headlessly, scored with a fixed rubric. report.html is the visual grid, RESULTS.md the write-up, out/ the generated maps. |
talk/ |
Outline with speaker notes, the live deck.gl title map (title-map.html), brand assets, and build-grids.py, which feeds the slides' result grids from the experiment data. |
research/experiments/*.sh, *.py, *.mjs |
The harnesses. Model calls go through the claude CLI, the Codex CLI, a LiteLLM proxy and OpenRouter; keys are read from an untracked .env. |
Everything was researched, run and written with Claude Code in the two days before the talk; the session is linked from the commits.
The suggestions in the talk were turned into proposals the day before the talk, each disclosing AI assistance and revised after an independent review pass: draft PRs #10677 llms.txt and #10678 agent skill and AI guide against visgl/deck.gl, the AI-assisted contributions policy for the vis.gl developer process (visgl/tsc #20) with its deck.gl companion draft (#10680, dependent on #20), and the @deck.gl/json v2 proposal posted on the v2 tracker with the full draft RFC kept on the fork (the deck.gl PR for it, #10679, was closed in favour of the tracker where v2 incubates).
Code and harnesses: MIT. Slides and text: CC BY 4.0. Model outputs in results/ and mapbench/out/ are reproduced as generated for research purposes.