What I've built, what I haven't, and how I work.

The questions I already had were ranking questions.

Over a decade running food and beverage businesses in Bandung, then operations for a gym on contract — scheduling, trainer management, the forecasts that decide staffing. I moved into ML because the questions I already had were ranking questions.

The work on this site is that transition: a refresh-risk scoring model built during the FlyRank internship, a demand forecast that runs against real bookings, a lead scoring pipeline built under a zero-cost constraint.

The first version looked good until I checked which pages it was ranking. Most of the separation came from pages that could not decline at all — no prior impressions, so the outcome was settled before the model saw anything. Scored only on pages where decline was arithmetically possible, the ranking showed no advantage over picking at random.

No model of mine has run continuously in production.

• One has informed a real decision — a client ranking a gym manager used to set priority — but as advice, not as a system anyone depends on.
• No testimonial. The internship isn't finished, and the model has had one user, and it's me.
• No live demo.
• I've never trained anything at a scale where infrastructure was the hard part.

If you need someone who has shipped ML into a product team, that isn't me yet.

I look for the leak before I look for the score.

A number that survives checking is the only kind worth sending you. When I find a failure mode I can't fix, it goes on the failure modes page instead of quietly out of the report. I work in English, Indonesian, and Mandarin.

Before the subdomain for this site exists, I wrote out what happens when someone types the address — an explanation behind the scene, and the notes work as a checklist.

Send one dataset and one question you want ranked.