A11y Review Coach: Accessibility Review with Bob

THE PROBLEM Many websites are hard or impossible to use for people with disabilities: buttons that only work with a mouse, images with no description for screen readers, form errors shown only in red, text too faint to read. These issues slip through code review because most developers and reviewers are not trained to spot them. They are found late, by users, audits or complaints, when fixing them costs far more and people have already been excluded. OUR SOLUTION A11y Review Coach turns IBM Bob into an accessibility inspector that any team can drop into its repository. It is packaged as a reusable Bob skill: a procedure, a WCAG 2.2 AA checklist, a severity guide and a report template. A developer asks Bob to review a folder with the skill. Bob reviews each page in parallel using subagents, merges and de-duplicates the findings, explains every issue in plain English with the people affected and the exact code fix, ranks it High, Medium or Low, then fixes the High and Medium issues on request. Results are written both as a readable report and as data that feeds a live dashboard. TARGET USERS Front-end developers and code reviewers without accessibility training, and product teams who need to ship inclusive interfaces. A pull-request template and a 15-item checklist bring the review into the moment teams already pause to check code. HOW WE PROVED IT WORKS We built a sample library-booking website with planted accessibility issues and hid the answer key from Bob using .bobignore, so the review was a genuine blind test. We also planted two extra issues by hand that Bob never saw. Bob produced 29 findings (13 High, 14 Medium, 2 Low), detected [X] of the [22] planted issues, and fixed the High and Medium findings. Google Lighthouse, used as an independent check, measured the average accessibility score moving from [A] to [B]. WHAT MAKES IT DIFFERENT We report the limits honestly. Bob missed low-contrast text, including both hand-planted issues, because it reads code rather than seeing the rendered page, while Lighthouse catches contrast reliably. Bob, in turn, found contextual problems Lighthouse cannot judge, such as vague link text, broken focus handling in the booking dialog and a misleading heading structure. The result is a practical workflow: the AI reviewer and an automated scanner cover each other's blind spots, with a person verifying the result. The project was built by a non-developer using a QA/QC mindset: specification first, clear scope boundaries, blind inspection, independent testing, and a log of every non-conformance, including the agent's own mistakes.