🏛️ Architecture Portfolio AI

𝗦𝗶𝘁𝘂𝗮𝘁𝗶𝗼𝗻

An architect (FAUP, Porto) preparing an internship portfolio wanted 3D visualizations of her housing project, Habitação Plurifamiliar no Viso. She had partial 2D drawings: one street elevation, plans, two sections and a materials board. No model, no full set of facades.

Nothing off the shelf fits. Floor-plan-to-3D tools do single-floor interiors, and image-to-3D models fall apart on orthographic drawings.

𝗧𝗮𝘀𝗸

Rebuild the building faithfully from her drawings — her design is the source of truth — and deliver portfolio-grade output over a weekend: renders, massing views, a 3D model and a print portfolio.

𝗔𝗰𝘁𝗶𝗼𝗻

Measured the geometry straight from the vector PDFs and built the massing model in headless Blender. Floor heights and bay rhythm come from the section and elevation, not from guesses.

Conditioned image models on the clay render, the elevation and the materials board, and split the work by model strength: one for eye-level shots, one for whole-building geometry.

One agent cannot hold design fidelity on its own, so every render went through a tribunal: an independent panel of agents, including an adversary, scoring it against frozen acceptance criteria — massing, facade module, residential typology, continuity across shots, archviz quality.

Used a multimodal LLM loop to check each render against the drawings and fix what failed with targeted edits — current render as the only reference, one surgical instruction — instead of regenerating the whole run.

Same loop for the interiors, the print portfolio and the Three.js site.

𝗥𝗲𝘀𝘂𝗹𝘁

Six photoreal views, four massing views, a Blender/GLB model, a print portfolio and a live interactive site. Nothing shipped on a single model's judgment: each view cleared multi-model consensus against the drawings, with the AI assistance disclosed in the deliverable.