How the work gets made.

Three disciplines, one habit: treat the tool’s raw output as unfinished, and impose the corrections that get it to delivered quality.

AI Image Creation

ImagePost-production over prediction Stills · Nano Banana Pro · GPT Image 2

I treat generative image models as instruments with predictable weaknesses rather than finished-output machines. Nano Banana Pro and GPT Image 2 each ship with defaults that quietly degrade product work — an amber-warm grade, a studio-composite look, inflated scale on small objects, and over-smoothed edges.

The mechanics

Preventative promptingPrompts are built to correct defaults before they appear, with recurring corrections locked as hard rules.
Material precisionWhen scale has to be exact for canvas, acrylic glass, Alu-Dibond and textile ranges, I composite the product in post instead of trusting the model to hold proportion.

Objective

Images that read as photographed objects on a real wall or table, not as renders.

↑ Index01 / 03

UX / UI Creation

UXProduction-grade prototyping Shopify-replicable · WCAG 2.2 AA

I build prototypes that are meant to survive contact with a live shop, not sit in a slide. The work is iterative, Shopify-replicable, and reusable patterns are pulled into a shared design system so decisions carry across pages instead of being re-litigated each time.

The mechanics

Built-in complianceEvery page is worked against usability heuristics, WCAG 2.2 AA and platform conventions from the first pass. Accessibility — contrast, keyboard paths, touch targets, screen-reader order — is treated as part of the build, not a later audit.

Objective

Measuring against outcomes rather than taste. The Universal Configurator, for instance, delivered a 31% overall and 84% mobile conversion uplift over the previous tool.

↑ Index02 / 03

AI Video Creation

FilmSystematised production Seedance 2.0 · Kling 3.0 · Premiere Pro

I run video as a documented production system rather than a sequence of lucky generations. Everything is finished properly in Premiere Pro; the generated material is the raw stock, not the delivered cut.

The mechanics

Targeted model deploymentEach model does the job it is actually good at. Kling handles low-level product and gallery motion, while Seedance is reserved for character work, large-scale campaign film and lip-sync.
Failure loggingA running failure log covers the specific ways shots break — acting, thin-item geometry, camera timing, grade, audio — so the same mistake is not repeated across sessions.

Objective

Locked, written rules governing generation — from two-case lip-sync handling to positive-first phrasing.

↑ Index03 / 03

The work Fifteen cases — the methodology in practice Back ↗