Everybody uses
a black box.Nobody argues with it.
Ex-Heineken · AI adoption
It's all probabilities.
The model gives a range.
People want an answer.
I build what sits in between.
First person in, three times.
The first hire at a Chinese architecture firm’s European office, set up in Poland from zero. One of two founding members of MAD Architects’ parametric design group in Beijing. The first product person in Heineken’s Global Analytics & AI Hub. Each time, the hard part wasn’t building the thing. It was getting people to use it. The fourth time, it’s mine: Yaga Labs.
Nobody acts on a black-box answer.
Everybody uses AI for the small stuff. Nobody bets a real decision on an answer they can’t question. Four steps that worked at Heineken, and that I use with agents now.
Heineken spends €7.9 million on marketing and sales a day.
Someone decides where every one of those euros goes. The models were ready early. The people deciding still trusted their spreadsheets more.
Allocation AI decides where the budget goes. Promo Advisor decides what to promote. Smart Flighting decides when it runs. Then the measured lift ↺ decides next year’s budget.
Three tools from Heineken’s AI Hub, joined so each year’s results plan the next.
Allocation AI
Bayesian budget allocation across markets, brands and channels. Heineken reports that markets following it spend more profitably, with incremental gross profit in Mexico, Brazil and the Netherlands.
Promo Advisor
Diagnostics, scenario simulation and calendar optimisation, so every promotion is tested before it launches. Nine tails, nine scenarios.
Smart Flighting
Tests campaign timing against AI response curves, week by week. Replaced an Excel process with a simulator media teams plan in. The lynx waits, then pounces.
She complains about AI. Then she ships it.
Yaga Labs is my experimentation platform: agents, small AI tools and black boxes you can argue with. Pani Yaga is its conscience. She has opinions about your model, and she's usually right.
Back in China,
the fitness function was the eval.
Five years working in China, back and forth ever since. I wrote the scripts behind a facade and an opera house interior at MAD, and competition wins at AREP. A decade before "AI agent" was a phrase, I was building evolutionary solvers that bred building geometry from a fitness function.
Optimal was easy. Buildable was the job.
Harbin Opera House
My part: parametric timber interior, detailed down to fabrication.
Gangxia North hub
My part: parametric skylights and entrances for the competition win.
