Methodology note

How the AI works — and where it is simply arithmetic.

Most of what this industry labels AI is a rule engine. Rather than let you find that out in a demo, here is the honest breakdown, feature by feature: what is a fixed calculation over your own records, what is statistical self-correction, and what actually calls a language model. A technical superintendent should be able to interrogate every number on the screen — so nothing below is a black box.

Last reviewed: 12 August 2026. Operator: Georgios Zografos, Cyprus.

Deterministic rules

Predictive maintenance & remaining useful life

A fixed, auditable calculation over your own records: running hours against the maker interval, calendar due dates, condition readings and the job's own history. The calculation is fixed and identical for every fleet, auditable against your own records — the same inputs always produce the same score.

Limits: It is not a trained model and makes no claim to detect failure modes it has never been told about. It cannot see anything you have not recorded.

Deterministic rules

Fatigue & attention (human factor)

A deterministic index from timestamps of work already recorded in the system: how much fell in circadian night hours, task-session length, consecutive active days and open critical backlog. The calculation is fixed and identical for every fleet, auditable against your own records.

Limits: No biometrics, no wearables, no cameras, no keystroke monitoring — and none are planned. Below a minimum signal floor the engine returns no score at all rather than guessing, and says so on screen.

Deterministic rules

Cognitive load forecast

Projects planned work over the next 14 days against a modelled watch capacity, reduced as the measured fatigue index rises, and reports the daily overspill. Arithmetic on the plan, nothing more.

Limits: Capacity is a stated assumption, not a measurement of your crew.

Statistical calibration

Prediction accuracy & lead-time calibration

Past predictions are scored against what actually happened in your records, and the lead time or alert threshold for that class of warning is shifted by a bounded correction. This is per-workspace statistics over your own outcomes — a control loop, not machine learning.

Limits: Corrections are capped, auditable and reversible. Your data is never used to tune anything for another customer, and nothing is trained on it.

General-purpose language model

Regulatory digest, drafted text and the decision copilot

Prompts are sent to a third-party general-purpose language model (Google Gemini, and Anthropic Claude for longer reasoning) through a managed gateway, with your data filtered by our egress boundary first. Output is presented as a draft for a qualified person to accept, edit or reject.

Limits: These models can be wrong or invent detail. Nothing they produce is applied automatically: no job interval changes, no deferral, no certificate action, no email sent without a human pressing the button.

Deterministic rules

PSC risk, audit readiness and emissions figures

Published scoring rules and regulatory arithmetic (FuelEU, EU ETS, CII) applied to your entered figures. Every component of a score can be traced back to the record that produced it.

Limits: Regulatory penalty and pooling maths follow the published texts; they are a model, not legal or class advice.

Three rules we hold ourselves to

A human decides, always

Every consequential action — deferring work, changing an interval, closing a non-conformity, approving a recommendation — is taken by an authorised person and recorded in the append-only audit log with who, when and on what basis.

EU AI Act posture

We are an EU-based operator. Where a feature interacts with you as AI, it is labelled as AI in the interface. The rule-based engines above are described as calculations, not intelligence. Human-factor scoring is used for workload planning by the operator; it is not used to evaluate, rank or discipline individuals, and we do not sell it for that purpose.

Your data is not training data

Fleet and crew records are not used to train third-party models and are not pooled across customers. Calibration is computed inside your workspace, protected by row-level security.

On the phrase “self-learning”

There is no continuously trained neural network behind this product, and we no longer use “self-learning” as a headline claim. What exists is a closed feedback loop: predictions are recorded, scored against your outcomes and used to shift lead times within bounded limits. That is genuine self-correction, and it is worth having — but it is calibration, not learning, and calling it the latter would not survive a pointed question from an engineer.

See the calibration loop Security, continuity & support