Problem
Accessibility gaps were blocking revenue. Engineers didn't have bandwidth to fix the backlog, and the longer it sat, the more it compounded. I needed a way to unblock the work without waiting on engineering capacity.
What I did
I built a pipeline and got designers and engineers on board. Design ships intent packs, Claude Code generates the fixes and opens PRs with before/after screenshots, and Engineering reviews and merges. I streamlined Claude Code into the workflow so the a11y backlog could move without pulling engineers off their roadmap.
How AI was used, with guardrails
Claude Code runs the scans, generates diffs, and opens PRs. All the labor. But I built the pipeline so AI is structurally blocked from making authority decisions: no auto-merge, no unsupervised fixes, no AI-decided severity levels.
Where human judgment mattered
I designed the authority boundary so AI never makes judgment calls. Design defines the intent (heading hierarchy, label copy, announcement language), Engineering gates every merge, and ambiguous cases always route to a human. The pipeline enforces this structurally, not by trusting AI to know its limits.
Relevant detail
Won People's Choice at a company-wide hackathon in February 2026. I negotiated a 1–2 day PR review cadence with engineering leads, and every PR ships with before/after screenshots. Separate ticket tracks per product surface.
Authority Distribution
People's Choice at company-wide hackathon, February 2026. Proved the pipeline model in 48 hours.
Three-role separation: Design ships intent packs → AI generates diffs → Engineering gates merges.
Tier 1: auto-fixable (alt text, aria-labels). Tier 2: needs design input. Tier 3: architectural, always human.
1–2 day review loop negotiated with engineering leads. Before/after screenshots on every PR. Two product tracks running parallel.
Get in touch