Axe-core just made voice-control accessibility non-negotiable, Vercel's infrastructure is quietly becoming the AI backend of choice, and LangChain shows why evaluation design matters more than agent architecture when you're building health-critical AI. Today's posts connect production tooling decisions with regulatory realities.
Find Label in Name issues in every scan with Axe-core 4.14.0
Axe-core 4.14.0 promotes the WCAG 2.5.3 Label in Name rule from experimental to default, catching controls where the accessible name doesn't include the visible label text. This matters for voice control users who say "click Settings" but the button's aria-label is "User Preferences." The rule now handles word-by-word comparison, supports non-Latin scripts, and runs 20-40% faster (up to 15x on large pages), so there's no performance excuse to skip it.
// Failing: accessible name doesn't contain visible label
<button aria-label="User Preferences">Settings</button>
// Passing: accessible name includes visible label
<button aria-label="Settings - User Preferences">Settings</button>
<button>Settings</button>How Interfaze serves 50 million requests a month on Vercel
Interfaze runs deterministic AI models (OCR, document classification, extraction) entirely on Vercel's Fluid compute with zero separate infrastructure, handling 50 million monthly API requests. AI Gateway gives them automatic failover between providers without code changes, and preview deployments let them run month-long customer pilots of new model versions before merging. The pattern here is treating Vercel as a full backend runtime, not just a static site host, and using preview URLs as staging environments for model evaluation.
// AI Gateway config for multi-provider failover
export const runtime = 'edge';
export async function POST(req: Request) {
// Vercel AI Gateway handles provider routing
const response = await fetch('/api/ai-gateway/v1/chat', {
method: 'POST',
body: JSON.stringify({
model: 'gpt-4',
messages: await req.json()
})
});
return response;
}How Numan built an eval-first AI health coach with LangGraph and LangSmith
Numan discovered that standard single-turn or scripted evals fail for conversational AI in regulated environments, so they built per-category scoring on themed datasets and simulated full patient personas to test boundary cases. The breakthrough insight is that clinician-authored evaluation criteria drive safety more than agent architecture does, letting them deliver behavioral coaching while staying clear of medical device regulation. Their approach flips the typical LLM development pattern: instead of building the agent first and adding evals later, they defined regulatory boundaries as eval criteria, then built the agent to pass them.
Today's theme is production-grade defaults. Axe-core ships stricter rules without a performance penalty, Vercel's infrastructure absorbs AI workloads that used to need custom backends, and Numan treats evaluation criteria as the product spec. The common thread is tools that enforce quality boundaries before deployment, not after.