fullbuild.ai

fullbuild.ai · idea → design → engineering → audit ⟳ shipped

A portfolio built as an architect's working drawing set. Scrolling advances a set of four sheets, and the drawing builds itself as you go: the same building appears as a graphite sketch, then a dimensioned blueprint, then an isometric structural frame, and finally the poured, finished thing, lit only where work is actually live.

The four sheets

• STAGE 01 · IDEA (graphite). A loose orthographic elevation plots itself on load, stroke by stroke, like a pen plotter.
• STAGE 02 · DESIGN (cyanotype). The sketch resolves into a dimensioned blueprint stating the design intent: type family, ink budget, base grid, motion verbs, contrast floor. Night mode drops the sheet onto dark paper.
• STAGE 03 · ENGINEERING (concrete). The load-bearing frame, drawn as an isometric wireframe and quantified in a real structural takeoff.
• STAGE 04 · SHIPPED (revision red). The pour: the frame fills with material, and the one accent colour ignites on whatever is live in production right now.

Stack

Next.js (App Router) · TypeScript · React Three Fiber (3D walk-through, in progress) · GSAP + Lenis · Zustand · CSS Modules

Tailwind is deliberately absent. With no utility classes in the build, no gradient or blur utility can leak in and break the contract.

Develop

npm install
npm run dev        # http://localhost:3000
npm run build      # production build
npm run typecheck  # tsc --noEmit

The screenshots in this README come from the capture harness in scripts/capture.mjs: point it at a dev server on port 3117, add --dark for night mode. It freezes the GSAP ticker before each shot so the plotter, hinge, and pour land on stable frames.

Built through RTK

The quiet workhorse behind this repo is RTK, a Rust CLI proxy that sits between an AI agent and the terminal and strips the noise out of command output before the model ever reads it: git status, git diff, rg, test runs, all compressed to what actually matters. Same information, a fraction of the tokens, so every session spends its context on the work instead of on boilerplate.

The savings are measured, not estimated. Across 95,821 logged commands RTK has filtered 79.5% of output tokens, 19.7 million tokens and counting, with per-command before/after logs refreshed daily.

See the running total at savetokens.tips

License

MIT