Lab

Ideas & experiments.

These are my experiments in code and craft. I've been building on the web for 18 years and I'm not about to stop. Even as a design leader, I believe the only way to stay sharp is to stay close to the medium. This space is dedicated to hands-on exploration, primarily using AI to bring ideas to life.

Creativityis.life — Xbox 360 NXE dashboard recreation

Creativityis.life

Vanilla JavaScript, Three.js r128, SCSS, Web Audio API, Canvas API, Blender, Docker, Nginx

A retro-faithful, single-page recreation of the Xbox 360 New Xbox Experience dashboard, built as a personal site after my son bet me I couldn't pull it off. He lost.

  • Custom 3D avatar and Xbox 360 controller modeled, rigged, and animated in Blender, exported to .glb, holding 60fps in browser via Three.js.
  • Original Xbox 360 system audio (blade clicks, confirm and cancel tones, startup chime) wired to interactions through the Web Audio API.
  • Music player styled after the Xbox 360 music blade, powered by GTA Vice City radio stations and a Butterchurn visualizer.
  • CRT scanline shader, parallax floor, and a layered particle system that sit on top of every screen.
  • Single-page vanilla JS, no framework, no bundler. Roughly 4,400 lines in one file by design.
Live site
Dirty Sara-Soda — full brand identity for a mobile dirty soda business

Dirty Sara-Soda

Figma, Illustrator, Photoshop

Full brand identity and merchandising system for my wife's mobile dirty soda business in Sarasota, Florida. Built from scratch in five months, on the road in two.

  • Logo system and "Cup Guy" mascot with a custom walk cycle.
  • Full print collateral: business cards, postcards, loyalty cards, regular and seasonal menus, menu boards.
  • Branded merch line: shirts, hoodies, 5950 fitted hats, plus a 10x10 professional event tent, feather flags, and pull-up banners.
  • 16-foot trailer fully wrapped in the brand identity.
  • Approved vendor at Manatee and Sarasota County school systems and the Sarasota Polo Club. TikTok went viral inside the first two months.
Linktree Instagram TikTok Facebook
Job Search Ninja — Claude Code skill for automated job hunting

Job Search Ninja

Claude Code, Chrome MCP, custom skill manifest, Excel automation

A Claude Code skill that scans 16 remote job boards, verifies every listing in a real browser, scores fit against my profile, and auto-generates tailored resumes and cover letters for the strong matches. Built during my own job search.

  • Scrapes 16 boards including LinkedIn, Indeed, Glassdoor, Built In, Wellfound, RemoteOK, Himalayas, FlexJobs, We Work Remotely, and more.
  • Chrome-first verification: visits every job URL to confirm the listing is live and still accepting applications. No more dead listings padding the spreadsheet.
  • Fit scoring against my profile (Senior, Staff, Lead, or Director-level UX, $130K+, remote): 8-10 direct match, 7 strong fit, below 7 skipped.
  • Auto-generates a tailored resume and cover letter for any job rated 7 or higher, saved to its own per-company folder.
  • Color-coded spreadsheet output with tier headers and dedup by company plus title pair, so the same listing doesn't get added twice.
CI Watchdog — AI-powered code review and site audit

CI Watchdog

Claude API, GitHub, CircleCI, Playwright, Docker, Azure, Wrike

A two-bot AI proposal to automate code review and post-deploy site QA across ESAB's 27 sites, 35 countries, and 54,000 pages. Pitched, scoped, and shelved when Anthropic shipped native agents that ate the entire orchestration layer overnight.

  • Code Review Bot wired into the PR workflow, blocking P0 merges and warning on P1s with inline suggested fixes.
  • Site Audit Bot that walks the entire public site after each production deployment, looking for broken links, buttons, forms, videos, and visual regressions.
  • Build economics: $45K to implement and $1.6K/month to run, against a $200K and 9-12 month internal model train. Projected $120K annual savings, 28:1 ROI, 1.6-month payback.
  • Pre-defined kill criteria: 80% bug-detection floor, 30% false-positive ceiling, plus developer adoption signals.
  • Tight pilot scope: one 5-person team, 8 high-traffic sites, 3 months. Picked the highest-bug-rate team on purpose so the results couldn't be dismissed.

AI product design concepts

Claude API, LLM, Named Entity Recognition, BERT, Conditional GAN, Figma

Two product concepts developed during MIT xPRO's AI Product Design certification. No production code, no shipped features, just what happens when 18 years of UX collides with the first real look at how AI works under the hood.

  • Conversational Form. Replaces multi-step forms (67% abandonment industry-wide) with an LLM-driven chat that quietly extracts structured data via NER. Looks like iMessage on purpose: everyone already knows how messaging works.
  • ArchiveRestore. A conditional GAN trained on synthetically degraded historical photos to restore real archival damage at a fraction of the $200-per-photo professional cost. Ground-truth training data generated by reverse-aging digitized originals.
  • Responsible AI by default. Confidence scores always visible, low-confidence regions flagged for human review, originals never overwritten.