ML Ops Demand Sensing

Problem

The growth team needed better demand signals ahead of a North American retail expansion. Their existing spreadsheet model overreacted to promotions and lagged on regional launches. Analysts experimented in notebooks, but nothing shipped because there was no repeatable pipeline.

Approach

I treated the project as a data product:

• Partnered with category managers to enumerate the decisions forecasts needed to inform.
• Created a feature store in BigQuery and orchestrated it with Prefect, logging lineage for every transformation.
• Prototyped gradient boosting and Prophet baselines, settling on an explainable ensemble that business stakeholders could trust.
• Deployed the models to Vertex AI with scheduled retraining and automatic backtesting against holdout stores.
• Embedded insights inside Metabase dashboards with scenario toggles so planners could validate assumptions.

We documented every experiment, from hyperparameters to distribution shifts, and wrapped the inference service with observability hooks that paged the team when drift exceeded thresholds.

Outcome

Weekly demand forecast accuracy improved from 82% to 91%, and planners finally understood why predictions changed because feature attributions shipped with every forecast. The automated pipeline eliminated 70% of notebook handoffs, and the same framework now powers promo lift experiments. Demand sensing became an operational process, not an ad-hoc data science project.