AI-Driven Retail Execution Platform

Overview

The project involved delivering an enterprise data and AI platform for a multinational consumer-goods company to orchestrate daily sales-execution planning for its field teams across several major retail channels and international markets. The platform combines a medallion-architecture lakehouse on Databricks with a portfolio of production ML models that translate raw retailer feeds, inventory signals, compliance data, and third-party audits into a ranked set of outlet-level tasks delivered to reps each morning. The system operates as a multi-tenant codebase where each retailer channel is onboarded as a configurable tenant rather than a fork.

Achievements

Brought a full production ML portfolio (demand forecasting at multiple time horizons, behavioural segmentation, compliance scoring, stock-availability risk, pricing anomaly detection, and a final task-ranking model) online and into daily operation. Reduced onboarding time for new retail channels from a multi-month custom build to a configuration exercise. Established a fully typed, validated configuration stack that catches misconfigurations before pipeline execution, eliminating an entire class of runtime failures.

Responsibilities

• Architected the bronze/silver/gold lakehouse on Databricks with parallel bronze ingestion, dozens of silver transformation tables, and a downstream gold layer consumed by the ML pipeline.
• Designed and implemented the ML inference DAG with explicit task dependencies, combining gradient-boosted forecasting, unsupervised segmentation, rules-driven compliance flagging, and a final prioritisation step that blends model-impact scoring with recency/cooldown constraints.
• Built a schema-governance framework using typed column and table definitions for consistent DDL management and evolution across bronze, silver, and gold layers.
• Implemented tenant-specific variation points (data sources, engineered features, enabled/disabled model outputs, output schema) so that a single codebase serves all downstream channels without branching.
• Stood up the CI/CD pipeline with lint, strict type checking, security scanning, asset-bundle deployment to layered target environments, and automated semantic versioning.

Technologies Used