01 Baseline
Capture before work starts.
The tool records the visual state and accessibility tree before implementation.
01 UN accessibility project · AI reliability engineering
AI that doesn’t break things when fixing things.
An agent-oriented visual and accessibility-tree regression workflow that made implementation evidence part of the review.
The work.
Accessibility changes needed reliable checks against unintended effects elsewhere in the product. We built a regression tool and prescribed workflow around the way an implementation agent actually works.
The workflow captures a baseline at the start. After implementation and self-verification, it compares the new state with that baseline across both rendered visuals and the accessibility tree.
The result.
The reported outcome was a substantial reduction in unintended changes to unrelated features.
- Context
- UN accessibility project
- Service
- AI reliability engineering
- Evidence
- Visual and accessibility-tree diffs
- Review
- Agent inspection and human web UI
Prescribed workflow
Evidence travels with the change.
02 Diff
Compare after self-verification.
Once the agent implements and checks the change, it diffs the new state against the baseline.
03 Agent review
Inspect every relevant change.
The agent reviews the diffs, crops the useful evidence and includes it in the pull request.
04 Human review
Keep the evidence legible.
A web diff interface gives human reviewers a direct way to inspect the same visual changes.
Skills and a prescribed workflow turned the tool into a repeatable delivery practice, not an optional final check.
Make the workflow prove itself
Build reliability into the work.
We assess your AI-assisted delivery workflow and show you where it needs stronger checks, clearer evidence and safer handoffs. Explore our AI reliability engineering and web accessibility services.