Senior frontend engineer — React, TypeScript, design systems, five years shipping product.
Most of my code now gets written by AI agents. That only works if the standard doesn't move, so the work I publish here is about the part that stays human: deciding what to build, writing the contract an agent implements against, and building gates that a broken implementation actually fails.
cosmos — a browser-based 3D universe explorer: seamless zoom from intergalactic scale to planetary surfaces, real star catalogs (HYG, Gaia DR3, NASA Exoplanet Archive) blended with procedural content. React 19 · TypeScript · Three.js · Web Workers, in a 17-package monorepo. Proven against a ~4.7M-star Gaia pack through streamed octree LOD. Live demo →
The graphics are the hard problem; the engineering is the point. Three things in that repo transfer to any codebase:
- Gates that can fail. Every gate ships with a deliberately naive implementation that must fail it. A gate a broken version also passes is decoration.
- Measure before fixing. 30 root-cause writeups in
docs/research/, kept whether or not they confirmed the first hypothesis — including the ones where my first suspect was wrong. - Boundaries in the toolchain, not in a doc. Package APIs are frozen between milestones and enforced by lint, with error messages that tell you what to use instead. The ordinary design-system problem: consumers reach past your public API, and a rule that lives in a README erodes.
executable-specs — the method from cosmos, generalized: spec format, review doctrine, test philosophy, and the Claude Code skills that run it. Extracted from a real project, not a thought experiment.
Argentina · open to senior frontend and platform roles · LinkedIn · rossetmatias@gmail.com