Case studies.

Production-grade Playwright QA agent skill stack: strict POM architecture, durable locators, no-static-wait synchronization, production guards, and audit scripts. github.com/MrNewDelhi/qa-agent-skill

Pixel-level diff with semantic component grouping over Playwright screenshots. False-positive rate down from 31% to 4%. Includes an interactive in-browser diff demo and skeleton-loader stabilization writeup.

LLM skill that normalizes manual test-case drafts (title, preconditions, step grammar, tags) against a 1,891-case rulebook before publish to AIO. Downloadable .md skill bundled.

Gemini 3 pipeline: epic → linked tickets → video attachment → step extraction → Playwright MCP replay → RCA via Sentry + Victoria Logs + GitHub. Comment-only on prod. Adaptive: HAR file, solo ticket, description-only branches.

Multi-model Appium framework that re-identifies elements at runtime across DOM heuristics + vision models. Built at L&T for two T-Mobile US MVNEs. github.com/MrNewDelhi/self-healing-locator-framework

Open-source MCP server for Android Debug Bridge. 38 typed, Zod-validated tools across devices, shell, packages, files, screenshots, logs, intents, settings. LLM skill bundled. github.com/MrNewDelhi/adb-mcp

How I rolled out Codex's built-in browser to manual QA and non-tech teammates with a permissions posture security could live with. Always-ask, narrow allowed domains, short-lived sessions.

Two ChatGPT Enterprise bots on connectors / MCP / skills. One for triage. One for test-case generation. Identity-scoped, deployed cross-team.

MCP tool that wraps email-testing behind typed tools. Made specs and agents able to assert on OTPs / confirmations without per-spec polling glue. Later migrated to Gmail + Google Groups — same MCP surface.

Don't take a screenshot until the page is actually done. Wait-for-skeleton + layout-stable pattern that cut flaky visual tests by half on the Visual Diff Analyzer.

How I picked the models behind the QA agent stack. Browser-action accuracy, tool-call correctness, minimum tool calls, cost per task. Eight model families compared. Runtime hooks for the real cost log.

Headed Playwright via MCP with request-shaping injection. Kept production smoke verification viable without disabling upstream bot-detection.

Tiny macOS menu bar utility for keeping a Mac awake — only in the way you actually need. Display / system / full / timed modes. github.com/MrNewDelhi/caffeinate-menu-bar-widget

Production-grade Playwright QA agent skill stack with 23 sub-skills, hard contracts, audits, and a TS POM template.

MCP server for Android Debug Bridge. 38 typed tools, Zod-validated, with bundled LLM skill.

Multi-model Appium framework with runtime locator recovery.

macOS menu bar utility. Display / system / full / timed modes.

MCP tool wrapping Mailosaur email-testing API. OTP extraction, link extraction, inbox management — used in QA automation at Lotusflare.

Everything else — forks, experiments, and whatever shipped this week.