Case study
by Elena Sokolova · Machine Learning Engineer
Hybrid Prophet + LSTM forecasting API reducing retail stockouts 15% across 120 store locations.
Demand Forecast API powers inventory planning for a regional retailer with seasonal categories and promotional spikes. The product goal was a single internal API that planners could trust for weekly replenishment decisions.
Spreadsheet forecasts failed during promotions and weather events. Store managers over-ordered safety stock, tying up working capital, while high-velocity SKUs still stockouted during peak weekends.
Built a FastAPI service exposing forecasts at SKU-store-week granularity. Prophet captured seasonality and holidays; an LSTM layer learned short-term momentum from POS signals. PostgreSQL stored historical actuals and model versions with rollback support.
- 15% reduction in stockouts across pilot regions - Forecast MAPE improved from 28% to 19% - Planning team saved ~6 hours per week on manual spreadsheet merges
Real-time product recommendations lifting CTR 22% and revenue per session 14% for a 2M-SKU e-commerce catalog.
Role: Lead ML Engineer
Edge-deployed computer vision system achieving 96% defect detection accuracy on high-speed manufacturing lines.
Role: Lead ML Engineer
