Human-centric design.
Agent-scale execution.
2025–27
The design decisions stay with a person: who the building is for, who operates it, what it costs them. The execution — code, pipelines, computational tools — runs through AI agents working a plan I write, gate, and verify.
- Role
- Designer · Orchestrator · Verifier
- Agents
- Claude Code · Parallel subagents
- Stack
- Python · SQL · Grasshopper · HTML/CSS
- Outputs
- Apps · Tools · Pipelines · Design workflows
Design for the person who inherits it.
I came to architecture from operations, where I held the utility bill for a Silver-certified building constructed ninety degrees off its intended orientation, glass atrium facing full sun. Nobody who made that decision lived with it. The people who occupy, maintain, and pay for a building are the brief.
Run execution at agent scale.
I direct AI agents the way I ran construction: a written scope, bounded tasks, a report at each milestone, and an inspection before anything moves forward. It lets one designer carry software and data work alongside the design itself. Accountability does not transfer to the agent.
Define the user, the objective, and the constraints before any agent runs. An undefined objective is the most expensive failure mode in agentic work. Agents execute it quickly and confidently.
Break the work into subtasks small enough to verify. Each gets its own instructions, inputs, and acceptance check. Independent subtasks run in parallel.
Every step that can change the next one ends in a report and a stop. Destructive or irreversible steps wait for a person. Nothing goes live on an agent's word alone.
Check output against the source, not against the agent's summary of it. On this site, drafted copy that claimed Revit on SketchUp projects was caught and pulled for exactly that reason.
Agents start each session cold. Decisions, conventions, and open items live in versioned build notes and plans, so the next session resumes from the record instead of from memory.
bradmachado.com
This site. Rebuilt from a hosted template into hand-written static HTML and one stylesheet, delivered by Claude Code against a phased build plan with a checkpoint at the end of every phase.
Portfolio Generator
A Python pipeline that assembles the print portfolio from the live site: generator to self-contained HTML to headless Chromium PDF. 25 pages, 41 images, each used once, no text invented.
HVAC Rightsizing
Peak heating and cooling loads transferred from ResStock onto national-scale U.S. property records. Harmonization and stratification scripted in SQL and Python on Databricks with AI-assisted development.
Less Siteless
Models in the design loop: a Flux LoRA trained on site imagery, YOLO v.7 traffic analysis from CCTV recordings, Ladybug solar exposure, and Wallacei X optimization driving the form.
Artificial Context
Context generated from a fine-tuned image model, screened structurally in Karamba3D, evolved in Wallacei X, and fed to a MATLAB neural network (R = 0.93).
Agents make execution cheap. They do not make judgment cheap. The scarce skill is knowing what to build, for whom, and whether the output is true — and keeping a person accountable for all three.