Cloudbridge originated as a conceptual pedestrian bridge: a suspended grid spanning between two supports, conceived primarily as a structural system rather than as a formal object. First presented on Designboom in 2013 as a speculative proposal, the project gradually evolved into a testing ground for computational design methodologies.

In the years that followed, as artificial intelligence began permeating architecture largely as an image-generation tool, the project expanded into a critical exploration of AI’s substantive role in design. The central question became whether AI could meaningfully support informed decision-making, structural control, and the evolution of complex geometries, or whether it was merely generating visually compelling outcomes.

The opportunity presented by the Hong Kong–Shenzhen Bi-City Biennale of Urbanism\Architecture enabled the realization of a scaled physical model: a two-meter-long bridge prototype produced using SLA resin 3D printing. The model was engineered to perform as a real structure rigid, structurally efficient, and optimized to reduce weight and cross-sectional material.

At this scale, the primary challenge was to prevent mid-span deflection. A defining characteristic of Cloudbridge is the deliberate absence of bracing. Unlike conventional grid structures, which typically rely on diagonal bracing systems for stability, the bridge remains intentionally unbraced.

To ensure structural reliability, analysis tools including Karamba3D and Abaqus were employed. In parallel, the project extended its exploration into artificial intelligence by testing a Large Language Model, Google Gemini, as part of the broader investigation into AI-assisted structural reasoning and design evaluation.

In Cloudbridge the grid is not an image, but a system of points connected by elements, defined through coordinates, relationships, and data. Its geometry was described numerically, not through pixels, but through structured information.

Although the bridge exists as a scaled prototype, with forces and constraints differing from those acting on a full-scale structure, it provides a safe yet stimulating environment in which to test emerging design methodologies by establishing a controlled setting in which experimentation can occur without the risks associated with immediate real-world implementation.

Artificial intelligence, specifically a Large Language Model, was introduced in parallel with Karamba3D as an experimental reasoning and optimization layer. While Karamba3D performed deterministic structural analysis, the LLM operated on structured numerical data describing the grid, including node coordinates, connectivity, member forces, and boundary conditions. Rather than generating images, the AI was tasked with interpreting this data, reasoning about structural behavior, and proposing variable cross-sections and optimization strategies. Its outputs were frequently returned in structured formats such as CSV files, enabling direct comparison within the computational workflow.

Through this collaboration, AI was reframed as a tool that interprets architecture as data and relational logic, working alongside Karamba3D’s structural analysis instead of substituting it.  Within this workflow, Karamba3D functioned deliberately as a benchmark. Its validated analytical results provided a reliable reference against which the outputs generated by the LLM, Gemini, could be evaluated, compared, and stress-tested. This approach enabled the team to critically evaluate the AI’s reasoning processes, detect inconsistencies, and determine where its logic aligned with or diverged from established structural analysis methods.