“The model was not the deliverable. The window diagram was.”
Reviews
What operators said after sitting the grid.
These notes reference specific modules. One of them is lukewarm on purpose — glowing-only pages are not useful when you are choosing a lab.
The leakage audit in Module 3 caught a “days-to-renewal” field our warehouse had been quietly backfilling. Embarrassing, and exactly why we enrolled.
I still wish the live clinics ran later for Chiang Mai time zones. The prepaid drought features, though, are now in our weekly pack.
We came in convinced survival curves were an academic side quest. After Survival Curves for Product Teams, our trial-to-paid chart finally stopped treating pause events as instant deaths.
★★★★☆
Reviewed after Module 3 — Feature leakage audit. Clear, a bit severe, and slower than our internal bootcamp. The severity was the point. Would not skip it.
Telco Churn Signals (Thailand) was the first course that did not import US cable anecdotes into a prepaid market. SIM rotation examples matched what our fraud desk already suspected.
Forecast Floor office hours are useful. The sample warehouse notes assume more dbt fluency than our two-person CRM team actually has — we spent a week translating. Still better than another vendor demo.
Case study · streaming save desk
A regional streaming CRM team arrived with an XGBoost score already in production and a save-desk complaining that “high risk” accounts were often the ones already speaking to retention agents. During the Cohort Retention Lab they rebuilt the label to exclude in-flight save cases, removed a post-contact ticket count, and moved the decision window to T−14 against the billing anniversary rather than calendar month-end.
Eight weeks later the queue still overflowed on campaign weeks — capacity was never the lab’s to fix — but false urgency dropped enough that agents stopped ignoring the score. Siriporn’s team kept the monitoring sheet pinned next to the campaign calendar. They did not publish a vanity lift figure; they published a weekly disagreement log between model and desk.
Case study · prepaid drought without a data lake
A smaller MVNO analytics pair sat Telco Churn Signals (Thailand) with only weekly aggregates. The limitation was stated on day one: daily recharge graphs were fantasy. Instead they learned to treat multi-week top-up drought as a state, not a smooth recency decay, and to stop blending postpaid contract inertia into the same chart.
Their “model” at the end was still a rules layer plus a logistic baseline. That was acceptable. The win was a shared definition with finance for involuntary non-pay versus silent SIM expiry — a distinction that had been inflating churn slides for a year.
