Data foundations before AI

AI work falls apart when we skip basic telemetry. We lock these in before we talk models.

Event hygiene

We track the handful of events that define success: creation, activation, win, and fail. Each event gets a responsible team and a freshness promise so the model doesn’t drift on stale data.

Golden tables

A single, well-documented table for customers, sessions, and outcomes beats a dozen half-finished marts. We keep an ADR next to each table describing owners, refresh cadence, and grain.

Metrics that map to UX

Model metrics (precision, recall) only matter if they track to user outcomes. We couple them. For a lead scoring model, we track meetings booked, not just AUC. The simple formula we use for the blended score is:

$F_1 = \frac{2 \cdot (precision \cdot recall)}{precision + recall}$