06 · Personal & Academic Project
Warehouse Sales Intelligence
A full decision system built around validation before modeling
Business value
77% honest demand accuracy, a $271,070 seasonal opportunity, and a $462,768 November peak forecast.
01 · Business problem
307,000+ records contained a $7,800 anomaly, an 83% gap in 2018, leakage risks, uncertain seasonality, and insights locked inside technical tools.
02 · My approach
I built the project in layers: audit, Medallion architecture, leakage-safe modeling, statistical testing, forecasting, then conversational access.
03 · What I built
Bronze-Silver-Gold pipeline, DuckDB, XGBoost, Prophet, A/B testing, drift detection, and an AI agent with human approval.
What changed my thinking
The corrections, tradeoffsand decisions behind it.
- 01
The first accuracy was too good
Three leakage types inflated performance. Fixing them lowered the number but made it honest.
- 02
The summer spike needed proof
Statistical testing showed Beer drove 95% of the seasonal premium while Wine and Kegs behaved differently.
- 03
The AI layer came last
The agent became useful only after SQL, model, and forecast outputs were stable and reviewable.