learn

Articles and series

From the open rebuilds. Full method, full code.

series
5 parts prereq: Comfortable with logistic regression and A/B testing

Uplift modelling and incrementality

Why a churn score sends money to the wrong customers, and what to model instead. Builds to a working treatment-assignment loop with a permanent holdout.

  1. A risk model is not a decision model
  2. Two-model, class-transformation, and what breaks in each
  3. Designing the holdout you will be asked to give up
  4. Measuring uplift when the treatment is not random
  5. Putting an uplift model behind an approval gate
From the rebuild: Churn intervention loop
4 parts prereq: You have shipped an API and read a trace

Agents that touch production

Evals, human gates, and the failure modes that only appear once an agent can write to a system of record.

  1. What an eval suite is actually for
  2. Human gates: approval, veto, and the anchoring problem
  3. Making an agent refuse
  4. Cost, latency and the trace you will need at 3am
From the rebuild: Fraud review assistant
6 parts prereq: None — starts from the business decision

Forecasting for people who own inventory

Accuracy is not the goal. This series ends at a working-capital number, and treats forecast error as an input to an inventory policy.

  1. Your forecast is an input to a decision, not the output
  2. Hierarchies, reconciliation and why totals matter
  3. Promotions: the feature nobody owns
  4. Lead-time variance is the real problem
  5. From forecast error to working capital
  6. Running it weekly without a data team
From the rebuild: Demand forecast + reorder