Case study · Machine learning
When the honest answer is less machine learning
How one ranking model replaced a plan for 100, on a dataset of about 40 usable rows.
Facts
CASE / 01The engagement at a glance.
Intraedge Engineering Intelligence
Bridge and culvert replacement engineering
116 aligned records, about 40 usable, 38 raw columns
One ranking model plus deterministic engineering
Feasibility study, 11-class taxonomy, feature analysis, production architecture, data schema requirements, 9-week roadmap to a production pilot
The decision
Another bridge, or a culvert?
When a railway bridge nears the end of its life, engineers decide what replaces it: another bridge, or a culvert that is cheaper to build and maintain. Today that call comes from a hydraulic and hydrological study that ranks three replacement options, and it depends on senior engineering judgement.
The first plan
About 100 models on about 40 records.
The first plan on the table was to train about 100 models, one for every output field. With roughly 40 usable records, that many models would have memorised the data rather than learned from it.
The reframe
Only one thing needs learning.
Only one thing in the output needs learning: which type of structure fits. Everything else follows from engineering that already exists. Dimensions come from Manning's equation, AREMA clearance rules and standard sizing tables. Pass or fail checks come from published criteria.
So the problem became a ranking problem: pick the top three of 11 structure types, with no type repeated.
What the data showed
Rules first. Then a model.
Two rules do most of the work. The minimum effective opening at the 100-year flood separates culverts from bridges. Existing multi-span structures almost always need a bridge, whatever the hydraulics say.
Missing data was structured, not random and several columns were near-duplicates. Both shape what a model can be trusted with.
The architecture
DESIGN / NOT YET IN PRODUCTIONOne model.
Everything else computed.
Ranking model
A gradient-boosted ranking model (LightGBM) trained with a Plackett-Luce loss that weights the top pick most.
Engineering formulas
Engineering formulas for every dimension.
Confidence gate
Below 0.45 confidence, the case goes to an engineer for review.
Screening tool
A desktop screening tool for engineers to assess 5 to 10 bridges at a time, using USGS StreamStats flow data.
How it will be judged
Targets set before the build, not measured results.
Accuracy above 0.75
Target, validated by repeated stratified 5×5 cross-validation.
Recall above 0.90
Target, validated by repeated stratified 5×5 cross-validation.
Above 0.85
Target, validated by repeated stratified 5×5 cross-validation.
Where it stands
Next: a baseline model.
The feasibility study, architecture and roadmap are delivered. The next step is a baseline model. The most important input, minimum effective opening, lives only in the original study charts, so extracting it is the first job.
The plan to grow the data uses historical studies, simulation (HEC-RAS or HY-8) and the public National Bridge Inventory.
The lesson we bring to every project
01Find the one judgement an expert makes. Automate that. Compute the rest.
Have an expert decision you want to scale?
Start with a 2-week
feasibility sprint.
We find out if your data can answer the question, then build the model that does.
