“The identity-state map from week one is still taped next to my monitor. Before the lab I treated device_id and account_id as the same grain whenever the join was convenient.”
Arisa Phong, lifecycle analyst
Reviews
Notes from people who had to take the score back to work.
We keep the mixed comments. A lab that only attracts applause is usually selling atmosphere. Average written score across 201 reviews: 4.3.
★★★★☆
SEA Retention Clinic was specific on festival spikes. I wish the wallet-retry drill had used a larger sample; my own app behaves noisier than the extract.
Niran, Chiang Mai
Took Leakage & Validity because finance had started ignoring our champion model. Two hours in I understood why: we had been scoring with a feature that only exists after the user converts. Dry, slightly humiliating, exactly what we needed.
Anonymous client in ride-hailing
★★★★★
Pulse Ledger recordings are dense. I paused the taxonomy critique three times to rewrite our event dictionary. Not entertainment.
Helen Cho · Ho Chi Minh City
Case studies
Two longer accounts, lightly anonymised
Stopping a “churn” score that was really a promo detector
A Bangkok grocery app had a 14-day inactivity model that spiked every time a campaign ended. After the leakage week, the team removed campaign-exposure features that were only populated for users who had already opened the push.
They replaced the score with a hazard model on session gaps and a simple rule: do not message anyone whose last session fell during Songkran unless the gap exceeds their personal baseline. CRM volume dropped; complaint tickets about “why are you chasing me” dropped with it.
The reservation they still hold: calibration plots looked acceptable in-lab and drifted after a wallet-partner outage. They now re-fit quarterly, which the syllabus only mentions, not drills at length.
Making LTV talk to collections without inventing certainty
An alumni circle member brought a probabilistic LTV notebook that collections wanted to treat as a credit rank. Over four critiques we forced the memo to say what the model cannot see: off-app bank transfers, family-shared devices, and agents who complete onboarding for the user.
The operating note now lists three uses (nurture timing, not limit increases) and one forbidden use. That constraint was more valuable than another decimal of AUC.