I build fast, scalable interfaces and the AI workflow that ships them.Ibuildfast,scalableinterfacesand the AI workflow that ships them.

case 01 · 210 sites

Problem

Thousands of small businesses have a website that is broken, invisible to search, or twelve years old. Nobody is going to tell them, and none of them are going to find me.

What I built

A crawler that scores their existing site across about thirty checks, a generative engine that designs each site from that business's own story, automated Cloudflare deploys, and an operations dashboard.

The AI's role

Agents research each business and write its site. The model never touches the outgoing message: that path is deterministic, because a sentence a stranger reads has to be auditable before it is clever.

Result

568 leads scouted, 210 sites built across 14 verticals, 177 live.

case 02 · no two alike

Problem

Ask a model for two hundred bespoke designs and you get two hundred variations of the same one. Every site being unique is the entire product promise, so convergence is not a quality issue, it is a refund.

What I built

A fingerprint registry. Every design declares eight axes: layout archetype, hero pattern, type pairing, palette family, texture, imagery treatment, price-list presentation, micro-motion.

The AI's role

The model proposes the design; the registry vetoes it. A new theme may share at most one axis with any single existing entry, and the build gate refuses a theme whose entry is missing or dishonest.

Result

210 shipped sites and the veto still has room. It also caught two CSS prefix collisions that were silently rendering one design with another's colors.

case 03 · 80+ built unattended

Problem

Building a site is an evening. Building at fleet scale means the evenings have to happen without me.

What I built

A scheduled overnight workflow that picks the businesses, builds and deploys their sites, verifies each one live, writes the outcome back into 14 vertical pipelines and two boards, and leaves a machine-readable summary.

The AI's role

A supervised multi-agent workflow doing the whole job unattended, inside a window that never races me, holding a lock so two runs cannot overlap.

Result

80+ sites built with nobody watching, over 15 nightly runs; best night 9 of 9. It aborts on preflight rather than half-finishing, which is the behavior I actually wanted.

case 01 · carrier-grade telecom

Problem

Business analysts at carrier-grade telecom operators need to configure data pipelines, reporting and real-time event processing without touching code, and the portal took 1.2 minutes to load.

What I built

I own the Data Hub frontend on the DNO Cloud Portal: data-dense configuration UIs, a step-based editor flow, and a multi-file S3 upload with per-file progress and DLQ recovery.

The AI's role

The AI work here is the workflow, not the product: I architected the team's AI engineering stack - cross-IDE agent rules, a context system, custom skills - and the team develops with it.

Result

Load time down 80% to 14 seconds. A 2,000-line editor refactored to 45% less complexity with zero regressions. Led the Vue 3 + Pinia migration; composables adopted team-wide.

case 02 · shipped solo

Problem

Board-game scoring is paper, phones die, and nobody keeps a group's history.

What I built

An offline-first PWA that needs no account and no server: flexible scoring, 20 curated templates, shareable Victory Cards, play-group stats, and a paid Pro tier with licensing.

The AI's role

Designed, built and shipped solo with Claude Code as the force multiplier, including the payments and licensing path.

Result

Live at gamescore.cards, with a paid Pro tier. Design to deployment, one person.

case 03 · data-dense UI

Problem

An analytics product with a data table that had to filter, sort, edit inline and select rows, over charts that had to stay fast.

What I built

The analytics dashboard and the DataTable, plus the migration from Redux Toolkit to TanStack Query.

The AI's role

None - this one predates the workflow, and it is here because the frontend half of the job is not all AI-assisted.

Result

Build time down 40% (18.8 → 11.4 s) and the main bundle down 53% (1,834 → 858 kB) via route-based code splitting.