ML forecasting systems with Endeavor Labs
Endeavor Labs brought us in on work for Central Storage & Warehouse, a temperature-controlled warehouse provider. The warehouse needed to plan labor around its customers’ production and delivery schedules, which shift constantly and don’t wait for the staffing rota to catch up.
Get the forecast wrong in one direction and you pay for idle hours. Get it wrong in the other and shipments go out late. Either way the cost is real, and it lands on the people planning the next shift.
The Signals Were There, in the Wrong Shape
Picking and packing throughput depends on how many people are on shift, which production schedules the facility’s customers have published, when deliveries are due, how demand moves with the seasons, and what the location’s own history says about all of that.
None of those signals were missing. They lived in operational systems built for tracking what happened, not predicting what comes next. Our job was to turn those records into a model that planning decisions could actually rest on.
Turning Records Into Features
We started with feature engineering, because a model can only weigh what it’s given in a usable form. Production schedules, delivery behavior, and seasonal demand each became structured inputs the model could weight, rather than raw records it had to make sense of on its own.
From there we built custom ML models that predict labor needs from historical warehouse data.
Choosing a Model That Holds Up
We used CatBoost gradient boosting classifiers. Gradient boosting is a strong fit for tabular operational records like these. We tuned the models with hyperparameter search and validated them with cross-validation, so performance holds outside the training window rather than only on the data the model has already seen.
The outcome was a production-ready planning input, not an offline experiment.
What Changed for the Warehouse
The forecasting work drove a 15% improvement in picking and packing efficiency per person-hour. It also gave the temperature-controlled warehouse services a data-driven basis for operational planning, with production-ready CatBoost models built on feature engineering and hyperparameter optimization.
The engagement also included collaboration on an NLP project, where we supported model architecture, training dynamics, and data preprocessing pipelines. It’s the same applied-ML engineering the forecasting work depends on.
The stack was CatBoost, gradient boosting, feature engineering, hyperparameter tuning, Python, and scikit-learn.