Event taxonomies before models
A working order for naming, versioning, and killing events so a classifier is not trained on folklore.
Journal
Field notes, not recaps of other people’s slides.
Short essays we use as pre-reading in labs. They assume you already have an event stream and a stakeholder who wants a number by Friday.
Retention signals we keep seeing in Thai consumer apps
PromptPay retries, LINE hand-offs, and the drop after a festival campaign — patterns from Bangkok product teams.
When churn models lie: leakage in mobile cohorts
The joins that look into next week, and why a beautiful AUC can still be a time-travel trick.
Probabilistic LTV without waiting a year
What a seven- or thirty-day window can honestly support, and the sentences you should refuse to write.
Reading session gaps as a leading indicator
Personal baselines versus calendar shocks, and why “days since last open” is rarely enough on its own.