Physics plus a learned correction · the incumbent, reimplemented

Warehouse Pallet Estimation Model

A hybrid: a physical packing model carries the structure, and a learned term absorbs what the physics cannot see, with each separately inspectable. The methodological decision I would defend hardest is that I implemented the incumbent rule of thumb in code so the comparison was apples to apples rather than against a remembered number, plus a detailed single-site pilot, because an aggregate improvement can hide a site where the model is worse.

2025mldata

What I owned

Owner. Packing model, estimation, aggregation, learned correction, baseline replication and pilot evaluation.

Mean absolute percentage error: 31.9% → 6.0%observedMean absolute percentage error
Total error: 15.2% → −0.3%observedTotal error
Carton records behind the packing model: 3.9McodeCarton records behind the packing model

2decisions, argued

Built onPython modelling stackA code replication of the manual process

01Problem 02My role 03Decisions 04Impact

The problem

What was actually wrong

Hub space planning used a commodity-level per-pallet rule of thumb. It ignored the physical differences between parts, so it was wrong in both directions and consistently wrong in the same places.

The information needed to do better already existed as carton records: dimensions and volumes per unit, across tens of thousands of part numbers.

Before this existed

A commodity-level dollars-per-pallet rule of thumb applied uniformly.

Constraints

The conditions the design had to hold under

Constraints are the interesting part of an architecture. Without them any diagram looks reasonable.

The incumbent was a remembered rule, not a specification: Comparing against it required implementing it, because there was no runnable version to compare against.
Physics does not explain everything: Actual pallet utilisation reflects handling practice, mixed loads and site-specific behaviour that no packing model captures.

My role

Owner. Packing model, estimation, aggregation, learned correction, baseline replication and pilot evaluation.

Designed

The three-stage hybrid: physical packing, part-level estimation, hub aggregation with a learned correction

Built

The packing model from millions of carton records, the estimation pipeline, the correction fitted on thirty-one months of actuals, and a code replication of the incumbent rule

Decisions

The calls I would defend

Each one with the alternatives I rejected, what the choice cost, and how it turned out.

Decision

The existing manual process was replicated as a pipeline stage so both methods run on the same inputs.

Context

The incumbent was a practice, not a system, so there was nothing to compare against directly.

Alternatives, and why not

Compare against remembered or reported figures for the old methodDifferent inputs, different periods, and a comparison nobody can reproduce. It is the easiest way to produce an improvement that evaporates under scrutiny.

Rationale

An improvement claim is a comparison. If the thing being compared against cannot be run, the claim cannot be checked.

What it cost

Building a version of something you intend to replace.

Outcome

Mean absolute percentage error of 31.9% for the replicated incumbent against 6.0% for the model, on identical inputs.

Impact

What changed, and how it is known

Every figure carries its basis. Nothing here is rounded up, and nothing modelled is presented as a result.

31.9% → 6.0%observed

Mean absolute percentage errorModel against a code replication of the incumbent rule of thumb, on identical inputs.

15.2% → −0.3%observed

Total error

3.9Mcode

Carton records behind the packing modelCovering 34,835 part numbers.

31 monthscode

Actuals used to fit the correction

Business: Warehouse space planning rests on the physical characteristics of what is being stored rather than on a uniform rule of thumb.
Engineering: A hybrid model whose two components fail differently and can be inspected separately, validated against a runnable replication of the process it replaces.

Leadership and hindsight

What I influenced, and what I would change

Technical leadership

Built a single-site pilot evaluation as well as the aggregate comparison, because an aggregate improvement can conceal a site where the new method is worse.

What I would do differently

The learned correction is fitted once. Handling practice drifts, and there is no monitoring that would tell anyone when the correction has gone stale.
Coverage of part numbers with sparse carton records is weaker than the headline error suggests, and the error figure should have been reported by coverage tier rather than only in aggregate.

Stack

What it is built on, and why that

A technology list without reasons is a list of things I have heard of.

Python modelling stack: The packing model, estimation, aggregation and the learned correction.Tabular problem, modest data volume, and the value is in the decomposition rather than the algorithm.
A code replication of the manual process: Baseline comparison.The only way to make an improvement claim reproducible when the incumbent is a practice rather than a system.

3 more sections are written and hidden: constraints, leadership, stack.