Teaching a transformer to think like your best planner.
A multi-station line-maintenance provider wanted to know why some turnarounds ran twenty minutes over — and others didn't — on the same aircraft type, every time. We trained a transformer on their own process history to find out, then distilled it into a model light enough to run at the gate.
A process is just a sequence with branches
Turnarounds, approvals, task-card releases — aviation processes are sequences of steps where context matters: what happened three steps ago changes what should happen next. That's exactly the pattern transformer architectures are built to capture.
Off-the-shelf process-mining tools flatten a process into a single "happy path" and treat everything else as noise. We treat the branches — deferred defects, MEL invocations, unscheduled task insertions — as part of the signal, not exceptions to discard.
So we trained a custom transformer directly on the client's own process logs: every step, every branch, every timestamp, for every station.
Teacher model learns everything. Student model runs everywhere
The full model — the "teacher" — is large enough to capture every nuance of the client's process history. We then distil that knowledge into a compact "student" model that's fast enough to run inline, at the gate or the workshop, without waiting on a data centre.
Most of the process is routine. The risk lives in the branches
The model maps the process as a graph: a core loop that runs the same way most of the time, and the branches that peel off it — some are genuine risk, some are hidden opportunity.
How the model works, step by step
Every process-optimisation engagement follows the same five stages — whether the process runs on the ramp, in the workshop, or in an approval queue.
-
We ingest your processes.
Every process log, every timestamp, every branch — ingested directly from the systems you already use, no manual tagging required.
-
We deep-analyse the core loop and its branches.
The transformer maps the process as a graph, separating the routine core loop from the branches that peel off it.
-
We look for risk and opportunity.
Each branch is scored: some represent genuine operational risk, others are hidden efficiency gains waiting to be captured.
-
We advise on the optimised process.
Concrete, sequenced recommendations — not a generic best-practice template, but a process redesigned around how your teams actually work.
-
We recommend KPIs and implementation discipline.
The optimised process only holds if it's measured. We define the KPIs and the review cadence that keep it that way.