Intelligent Markdown Engine
Decides which products to discount and by how much, leaving the decision ready to load into the company's ERP.
The problem
Markdowns were decided by hand, by intuition, across 11-country catalogs: the same discount for everything, giving away profit on products that did not need it.
What I did
Designed and built the entire engine: the full markdown decision logic —rule-based eligibility, sale-price calculation, per-country margin control and product prioritization—, a GPT-4o executive summary and help agent, and direct export to the corporate ERP format. Evolved it in production through six months of continuous releases.
The outcome
In production as the approved channel for loading markdowns into the ERP, and it grew five-fold in scope over six months.
How it decides
- 11-country catalogInput
- Rule-based eligibilityWhich SKUs may go on sale
- Sale-price calculationHow much each product drops
- Margin controlPer country
- Assortment prioritizationWhat drops first
- Export to the corporate ERPExecutable output
Each step is a codified business rule, not a suggestion: the engine outputs a decision ready to execute. An LLM writes the executive summary over that output, never over the raw data.
Stack
- Python
- Streamlit
- SQL Server
- LangChain
- GPT-4o
- Plotly
- openpyxl
- Docker
Blue Star Group internal project — the code is not public.
Next project: Seasonal Buy Optimizer →