AI / ML — Sep 14, 2024
在未来到来之前看见它。 See it before it arrives.
Trained and deployed a forecasting model that improved inventory planning accuracy by 28% quarter-over-quarter. Fourteen data sources unified into a single signal.
The data problem
Fourteen data sources, zero shared schema. Before any model work we spent three weeks building a normalisation pipeline that could handle late, missing, and contradictory signals.
パターン認識 — Pattern recognition
The model learned to weight recency over volume — a Friday in November means something different from a Friday in April. Seasonal priors, encoded explicitly, cut MAPE by 11 points.
Shipping fast, failing faster
We deployed to 5% of SKUs first. Watched for three weeks. Expanded to 40%. Watched for two. Full rollout at week seven. Every metric was instrumented before we wrote a single line of model code.
モデルは仮説だ — Models are hypotheses
A model is only as good as your willingness to be wrong about it. We shipped weekly and killed features that did not move the metric. Certainty is the enemy of accuracy.
← Back to projects