All projects
Apature
VLM design review that checks judgment, not pixels
Startup · Founder · June 2026 — Present
Overview
Apature is my design tooling startup. It builds automated design review for rendered UI, split across six public MIT repositories that connect through shared wire contracts. I wrote nearly every commit and published five of the packages to npm under @apatureai.
Verdict
The core product. One command measures a running page in headless Chromium (WCAG contrast, overflow, touch target size) with no API key. Add a vision model and it critiques the screenshots against the repo's own design system, then deletes every finding it cannot point at a captured element.
Gate
A GitHub Action and GitHub App that run a pull request's preview build in a hardened sandbox, send the preview URL to a critique service, and post the review as one sticky comment plus a Check Run. It never requests write access to code and never shows a passing review for a page nothing judged.
Sigil
Error bars for LLM-as-judge evals, in dependency-free TypeScript. Given judge verdicts and human labels, it reports calibration, finite-sample risk certificates, and Pass^k reliability, and only recommends a cheaper model when the evidence supports it. Every report is byte reproducible.
Bastion
A remote MCP server that gives a coding agent eyes on the UI it just changed: submit a preview URL, get findings with element references and suggested fixes, then ask for a recheck. It doubles as a reference for OAuth 2.1, SSRF-safe URL handling, and long-running jobs over MCP, and is listed in awesome-mcp-servers.
Lattice
A deterministic scene graph for browser agents. It fuses DOM, accessibility, computed style, and text capture into one content-addressed snapshot, keeps both claims when sources disagree, and renders small token-budgeted views for a prompt.
Canon
A strict DTCG 2025.10 design token resolver and project scanner for CSS custom properties, Tailwind, and token files. Broken or circular tokens are reported as diagnostics instead of guessed, and a 20-repo conformance corpus measures how much published token data passes the strict profile.