Case Studies
The details are anonymized, but the constraints were real: messy data, tight timelines, and stakeholders who care about outcomes more than models.
Built an end-to-end forecasting pipeline across banners, channels, and store clusters using a mix of classical time-series models and tree-based methods. Introduced strict backtesting, scenario overlays for planners, and automated monitoring for drift and data issues.
Translated planners’ allocation rules, pack logic, and carry-over behavior into a parameterized engine. The system consumed demand forecasts, pack libraries, store attributes, and DC constraints, generating proposed flows that planners could review and override with full traceability.
Designed a metric layer and dashboards on top of the enterprise data warehouse, then added an LLM-powered assistant that generated performance narratives, highlighted anomalies, and answered ad-hoc questions using governed, audited data rather than raw exports.