Co-authored with Claude. Final editorial judgment and accountability are Alex's.
I'm really excited to speak at Drupal GovCon this year — August 11–13 at the University of Maryland's Stamp Student Union in College Park. We've been a part of the Drupal4Gov community for years and we've attended and sponsored both this event and the previous Drupal4Gov versions. The event is free to attend, and it truly is one of the most consistently well run and informative conferences around, so if you're in the area I hope you'll join (it's not just for the "govies" either — any open source developer should consider coming).
The talk: Proof, Not Vibes
My session is called "Proof, Not Vibes: AI-Assisted Accessibility for the Drupal Sites You Run" — Wednesday, August 12 at 9am. The TLDR version: AI can find and fix accessibility problems on your sites — if you provide it the right AI skills and tools and you check its work. AI output can be deceptive. A patch can look done while a keyboard user still can't finish the task. The talk is about the tools that can do the work as well as the checking (much of it automated).
The Accessibility Skills toolkit
Everything in the talk comes from our open source (GPL) Accessibility Skills toolkit, which you can find on GitHub. It's nine skills that run in Claude Code, Codex, or fully locally through Ollama:
- Plan and critique. A planner that makes accessibility decisions on purpose, before code hardens them, and a critic that reviews plans and implementations — where severity means "how blocked is a person," not "how important does the rule sound."
- Audit from every angle. A perspective audit built on CivicActions' personas (who is blocked?) and a role audit based on W3C's ARRM (who on the team owns the fix?).
- Test with real evidence. Real keyboard events in real browsers via Playwright and axe-core, journey audits of live URLs, and screen reader assertions that run in CI. Never synthetic key events, never attribute-checking dressed up as testing.
- Findings that survive scrutiny. Every finding carries the element, the WCAG criterion, reproduction steps, and who it blocks — and turns into a reproducible bug report (adapted from Mike Gifford's open guidance).
- Benchmarks you can read. The same test fixtures run across Claude, OpenAI, Gemini, and local models, with raw results committed to the repo. Mid-July we caught our own benchmark quietly helping the models pass. We fixed it and published the lower numbers.
[Session slides will be added here after the talk.]
Come say hi
If you're at the event I hope you'll consider coming to my session on Wednesday at 9am or stop by the Zivtech booth and say "hi!"
If you're not at the event, keep an eye out here for the slides, and reach out if you want to chat about AI-enabled accessibility remediation, something we specialize in at Zivtech — planning, testing, fixing, and proving it. We just launched a new accessibility services page with the details.
If you'd like an accessibility assessment of your site — no cost, no obligation — I'd love to do one. Find me at GovCon, or get in touch at alex@zivtech.com.