Design, Implementation, and Deployment of a Machine Learning Framework for Real-Time Packaging Optimization in Industrial Logistics.

Now, packaging decisions at AMA S.p.A.’s logistics plant are 100% a matter of human choice: when an order is transmitted by an automatic sorter to a loading bay, the operator has only a few seconds in which to decide which boxes to use and how many. This is an example of the bin packing problem, a computationally hard problem that human operators are required to solve near-instantaneously for each order of a shift, leading to predictable inefficiencies – mainly over-packaging, used as a hedge to the cost of an undersized choice, and performance variability between experienced and inexperienced staff. Instead of trying to replace this decision, this thesis designs, implements and deploys internally an end-to-end machine learning system that supports this decision. The system is formulated as an imitation-learning problem, predicting from an order’s composition what box types are needed and in what quantities. The system learns from AMA’s own historical packaging decisions, rather than from item geometry directly – a decision mandated by two constraints set during requirements analysis: CPU-only inference hardware and article dimension data of questionable reliability. The model employs a two-stage classify-then-regress architecture, a multi-output Random Forest followed by per-class conditional regressors, trained on approximately 110,000 historical orders. The surrounding system is constructed as production-oriented software instead of a research prototype, with a containerized, six-service architecture, versioned datasets and models, a REST API to connect the system to the existing Warehouse Management System, and a shadow-mode deployment that logs live predictions without exposing operators to unvalidated suggestions. The model does not reach deployable performance by its own pre-specified metrics: support-weighted precision on classification is still around 0.30 and the regression stage yields a near-zero global R², with two severely negative outliers hidden within. The thesis does not take this as a final outcome, but rather traces the outcome to its origin by directly auditing the data. It finds that roughly 14% of the catalog of articles contains placeholder dimensions -- syntactically correct but physically meaningless values that pass all structural checks and propagate undetected through the entire pipeline. This condition is called operational blindness in this thesis, a systemic defect that makes the physical reality of an order invisible to the model and indeed to any system or person who relies on the same record. The thesis argues that the main contribution of the project is diagnostic rather than purely predictive, by separating this cause from two other reasons for the reported metrics, expected fallback on data-scarce box types and one still-unexplained anomaly: a quantified evidence-based account of why the data could not yet support the system it was meant to train, along with the reproducible infrastructure and monthly sanitation roadmap to close that gap.