Proof of work

Case studies

Real builds for real businesses, with the receipts: recorded flows, honest numbers, and what each system actually does.

Animarek · Tool trust

I tested if the new meeting note taker from WisprFlow is to be trusted in my pipeline. It isn’t.

The transcript read beautifully and put the wrong name in front of half of it. Two files and some counting exposed a cliff no amount of reading would have caught. The three-move check you can run on any tool before it goes under your business.

8,039 aligned anchors · 84.2% overall, 6.4% in the final five minutesRead the case study →

Animarek · The handwriting-to-post pipeline

Write by hand. Publish like a studio.

Every morning I write with a pen, on paper, on a timer. Minutes later the pages are digitized word for word, and while I do my one manual edit, my own character is already illustrating the post. The AI never writes a sentence.

18 consecutive handwritten posts · digitized, illustrated, publishedRead the case study →

Anitinke · Visual identity system

One character. Locked once. Rendered anywhere.

How my character Anitinke stays recognizably the same across hundreds of AI-generated scenes: lock a canon once, then every image is references plus one instruction. A face for the projects that were never going to have one.

376 cataloged images · 8 scenes, one character · built from the live catalogRead the case study →

La Hora Azul · Custom point of sale

Your business runs on its own rules. Generic software doesn't understand them.

A consignment art shop where every piece belongs to a different artist, each with their own commission. Commissions compute themselves, discounts recompute the split, balances stay current. Built in one day.

7 recorded flows · built in 1 day · live on VercelRead the case study →

X-Factor · Take selection

Seven takes of the same line. Only one makes the cut.

A tool that recognizes every repeated take in a recording, groups them, and makes picking the best one simple. The creative call stays 100% yours.

every take in a 9.6-min recording, picked in ~3 minRead the case study →