H-38.

Have a password?

Unlocks the client-named version of every project.

Site Data-to-Park Options

Landscape & urban design · Dubai · competition entry

A browser studio that pulls a real site's data, models it in 3D, generates scored park layouts and logs every AI step — built for a municipal design competition.

Re-enacted: the generated park layout with its shade heatmap and scored alternatives

Context

The brief carried a twist: the jury wanted an AI Methodology Report alongside the design, so every generated option had to be traceable to its inputs. The usual toolchain — GIS export, CAD, a rendering package, a chat window — produces a proposal but not a record.

  • Site context (buildings, heights, roads, footpaths) had to be assembled by hand from several sources
  • Alternatives were drawn one at a time, so there was never a fair comparison
  • Shade and heat — the design constraint that matters most in a 47 °C August — were checked late, by eye
  • Nothing recorded which decisions the AI had influenced and how

Outcome

6steps, locate to present

  • Site context went from a day of assembly to a fetch
  • Shade became a design input at generation time instead of a late check
  • PLACEHOLDER — add the competition outcome when known
Re-enacted: select the site, extract the surrounding city from open data, generate scored alternatives, and run the August shade analysis

What we built

A six-step workflow in the browser — Locate, Extract, Model, Generate, Evaluate, Present — where each step is a module registered in one place and consumes the previous step's artefact. Search a location or draw the plot; the tool pulls buildings, roads and paths from OpenStreetMap, infers building heights with provenance, and writes a layered CAD plan exportable as DXF.

A 3D context model with extruded buildings, a sun slider and shadows, and a SketchUp-style toolset — draw, push/pull, offset, section plane, tape, undo — so the team could sketch in the same place the analysis runs.

A seeded generative engine that produces four scored alternatives (a jogging loop, programme zones, a shade belt) from the site's constraints, and an evaluation step with a ray-cast shade heatmap, wind rose, annual solar radiation and sky-view factor from ERA5 climate data. Exports: PNG, DXF, a Markdown design report, and the project as JSON.

An append-only AI-usage log, written for the methodology report the competition required, plus a research pass that answered 45 questions about the brief with eight parallel research agents before design started.

What stuck

  • Real data first: every option sits on the actual site, not a drawing of it
  • Alternatives are generated together and scored the same way, so the comparison is honest
  • The AI log is a by-product of using the tool, not a report written afterwards

Snapshot

Client profile
Architecture and engineering team entering a municipal AI park-design challenge
Engagement
Catalyst — competition tool, then a feature roadmap
Timeline
Built 11–19 July 2026; roadmap locked 11 August
Scale
13 packages, ~12,000 lines, 5 architecture decisions, 45 research questions answered
Scope
Site location, GIS extraction, 3D context, generative layouts, shade and climate analysis, exports, AI-usage log
← All work