01
The problem
Operational teams were planning against forecasts that missed often enough to erode confidence in them. When a forecast is not trusted, people plan around it - which is more expensive than having no forecast at all.
02
What I did
- Built and maintained analytics environments the global network relied on for reporting
- Applied statistical modeling and machine learning to improve demand forecast accuracy
- Used experimentation and behavioral analysis to test whether changes actually improved decisions, not just metrics
- Translated model output into views operators could act on directly
Trade-off
Optimized for
Accepted cost
Reasoning
Model complexity vs interpretability
Models planners could reason about
Some headroom on raw error metrics
A forecast nobody understands is a forecast nobody plans against, so explainability was part of the accuracy story.
Metric gains vs decision change
Improvements that changed a planning decision
Wins that only moved an offline score
Experimentation was set up to test adoption and decision quality, not just model fit.
03
Outcome
Forecast accuracy improved by roughly 20%, and - as importantly - the outputs were adopted into how planning conversations were run.
A model that nobody plans against has zero accuracy in practice. Adoption is part of the quality bar.
04
Why this is the root of my AI work
Working directly with models taught me what variance looks like from the inside, how easy it is to over-trust an evaluation metric, and why the product question is always "which decision does this change?" That is the same question I ask of every AI system I evaluate now.