Rabia Yoruk

Kolay.AI demand forecasting

Forecasting for small and medium businesses. Models predicted demand swings and disruptions, and the clients used them to plan inventory.

Role
Data Scientist and ML Engineer
Where
OptiWisdom
When
2022 to 2023
Stack
  • Python
  • scikit-learn
  • OptiScorer
  • Docker
  • Kubernetes
  • CI/CD on AWS and Azure

82.5%

forecast accuracy

10

SME clients using the forecasts, approximately

Kolay.AI forecasting flow, simplified. Several models compete, the best one is kept. Flow: Sales data and Inventory data, then ETL, then Features, then Candidate models, then Model selection, then BI dashboard. Sales data client exports Inventory data client exports ETL clean, join Features per client Candidate models incl. OptiScorer Model selection compare, keep best BI dashboard inventory plans Kolay.AI forecasting flow, simplified. Several models compete, the best one is kept. Flow: Sales data and Inventory data, then ETL, then Features, then Candidate models, then Model selection, then BI dashboard. Sales data client exports Inventory data client exports ETL clean, join Features per client Candidate models incl. OptiScorer Model selection compare, keep best BI dashboard inventory plans
Fig. 1 Kolay.AI forecasting flow, simplified. Several models compete, the best one is kept.

Problem

Small businesses run out of stock, or buy too much, because demand changes and nobody sees it coming. They do not have a data team. They need a forecast they can read and act on.

What I built

  • Forecasting models that predicted demand swings and disruptions for about 10 SME clients.
  • Several candidate models, including OptiScorer, OptiWisdom’s licensed scoring model. They were compared on the same data and the best one was kept.
  • Customer segmentation with RFM analysis, used for recommendations and cross selling.
  • CI/CD for model versioning, retraining and deployment. Services ran in Docker on Kubernetes and AWS ECS.

The forecasts reached 82.5% accuracy and the clients used them to plan inventory.

Paper

The platform and its results are described in a sole-author paper in the Journal of Data Analytics and Artificial Intelligence Applications (2025).

50 100 150 200 w1w14w27w40w52
  • Actual demand
  • Forecast
  • 80% range
Fig. 2 Weekly demand for one product, with a supply disruption around week 32. Illustrative, synthetic data.