AI
Generative tools only pay off when their output is repeatable. I build the pipelines, prompt structures and plugins that turn lucky one-offs into a production process — and check the result against funnel data.
How I work with it
Each one came out of shipped client work, not a demo.
A five-stage concept-to-delivery pipeline that selects a different model for each shot type rather than forcing one tool to do everything — Midjourney, Kling, Runway, Seedance and Sora each pick up the work they are actually good at.
I publish the methodology bilingually on LinkedIn so a team can run it without me in the room.
JSON-structured prompting that separates camera, light, material and rhythm into their own fields. Because each variable is addressable on its own, a shot can be adjusted rather than re-rolled — which is what turns generative video from lucky one-offs into a stable production process.
The same structure is what makes brand-grade output possible: the look survives across shots instead of drifting.
A 19,020-node Figma file had drifted to 423 component masters. Cleaning it by hand was not realistic, so I built Component Doctor — a 2,222-line custom Figma plugin in vanilla JavaScript — and brought the file down to 27 canonical components.
Alongside it I designed the team’s AI creative pipeline and an 8-lesson Figma curriculum, so the cleanup held after it shipped.
I reverse-audited an agency’s 53-page paid-media report and surfaced the B2C-versus-dealer mismatch sitting behind roughly $45K of attribution, then redefined how conversion was tracked.
Diagnosing the GA4 and Shopify funnel turned up the break plainly: add-to-cart up 118% while checkout fell 74%. Design decisions after that were argued from the funnel, not from taste.
Receipts
Toolkit
Chosen per shot type and per problem, not by habit.