Howthy/Sample Build Record

AI & Data Capstone · BR-024

Capacity-Aware Predictive
Maintenance Triage

A decision-support workflow for prioritising vehicle inspections when maintenance capacity is limited.

Review ready

LAST UPDATED28 March 2026

VISIBILITYPublic sample

REVIEWFaculty-attested

01 / OVERVIEW

A constrained prioritisation problem.

A regional maintenance team can inspect only a subset of its vehicle fleet each week. The project explored whether historical operational data could support a relative risk ranking that helps human reviewers decide where to look first.

Problem

Which vehicles should receive limited inspection capacity first?

Success criterion

Higher concentration of relevant cases in the top-ranked inspection set.

Primary constraint

No verified physical-failure labels or component-level diagnostic truth.

Human role

Reviewers retain the final inspection and maintenance decision.

SUPPORTED CLAIMS

Relative risk rankingInspection prioritisationCapacity-aware decision supportReproducible evaluationHuman review

NOT SUPPORTED

Exact component diagnosisPhysical-failure probabilityRemaining useful lifeAutonomous maintenance decisionsGuaranteed commercial savings

02 / DECISION TRAIL

The choices that shaped the claim.

D-03
18 FEB · MODEL DESIGN

Define a ranking problem—not diagnosis

Vehicle-level labels and maintenance records did not support exact component diagnosis.

View linked evidence
D-08
04 MAR · MODEL DESIGN

Optimise capacity-aware utility

The operational question was which cases to inspect within a fixed weekly capacity.

View linked evidence
D-11
19 MAR · CLAIM GOVERNANCE

Keep commercial impact out of the claim

No production pilot or independently confirmed cost baseline was available.

View linked evidence

03 / EVIDENCE GRAPH

Claims stay connected to their support.

MATERIAL CLAIM · C-04Higher-risk cases are concentrated near the top of the ranking.Supported
NBEvaluation notebookSystem-captured · v1.8
TSTemporal split testTested · 3 windows
RVFaculty reviewAttested · 27 Mar

04 / EVALUATION

Tested as a ranking workflow.

Relevant cases captured by inspection capacityIllustrative sample data
10% capacity42%
20% capacity67%
30% capacity79%
40% capacity88%
Model rankingRandom baseline
KEY RESULT

The ranking concentrated more relevant cases within each tested capacity band than the comparison baseline.

This supports prioritisation utility in the retrospective dataset. It does not establish production performance, physical failure, or savings.

05 / AI-USE DECLARATION

AI assistance, in context.

Tools & purpose

Used for code explanation, test-case ideation, documentation structure, and wording review.

Human judgment retained

Problem framing, feature acceptance, evaluation design, claim boundaries, and final evidence selection.

Verification

Generated code was inspected, adapted, executed, and checked against held-out data and documented assumptions.

Rejected assistance

Suggestions that implied failure probability or commercial impact were rejected as unsupported.

06 / LIMITATIONS

What this project does not demonstrate.

!

Clear boundaries are part of the evidence.

The work has not been deployed in an operational maintenance setting. It does not demonstrate causal reduction in downtime or costs, nor validate the ranking against confirmed physical failures. The next responsible step is a time-bounded, human-supervised operational pilot.

RK
“The record makes a strong distinction between retrospective ranking performance and operational impact. The supported claim is appropriately bounded.”Dr Rana Khalid · Faculty reviewer · 27 March 2026

This is what evidence-backed work looks like

Build Records make the strength—and the limits—of project evidence visible.

Request a pilot