About · Team
AI and Software Team
This function builds the detection and data path, and is deliberately conservative about what it claims. A model that must be verified by a person should be presented as an aid, not as an authority.
Overview
- Models are evaluated on the client's own imagery before deployment; performance on a demo set is not a deployment result.
- Deployment is on-premise or customer-controlled by default, so mission imagery stays within the client's own boundary.
- Every automated output keeps its provenance: which model version, which inputs, and what the confidence meant at the time.
- A human review step is designed into the workflow, not added as a disclaimer.
What this function owns
The detection models, the data path that feeds and consumes them, and the deployment into whatever environment the client controls. It also owns the discipline of not overstating any of it.
The useful question in applied industrial AI is rarely whether a model can find something in a curated demo. It is whether the output survives contact with the client's own imagery, their own lighting, their own asset mix — and whether the person who has to act on it trusts it enough to act.
Evaluation on the client's data, before claiming anything
Models are evaluated against the client's own imagery before deployment, because performance measured on a demonstration set is not a deployment result. Where that evaluation has not yet been carried out, we do not publish a detection rate.
This is the reason no accuracy or detection-rate figure appears anywhere on this site. We hold many and publish none of them without the measurement record behind them, as set out in AI evaluation and integration.
Deployment, provenance and the human step
Default deployment is on-premise or in customer-controlled infrastructure, so mission imagery stays inside the client's own boundary. Where an exception is made, it is a specific architecture decision rather than a silent default.
Every automated output keeps its provenance: model version, inputs, and what the confidence value was intended to mean when it was produced. Without provenance, an output cannot be defended months later, and a confidence score without a defined meaning is decoration.
A human review step is designed into the workflow rather than appended as a disclaimer. Where a consequence is real, someone entitled to make the call makes it.
Related pages
- Technical StrengthView
- Company overviewView
- Development Testing and DeliveryView
- Custom Engineering NotesView
More in this area
- Team OverviewRead
- Leadership and Commercial LeadsRead
- Project Management TeamRead
- Systems Engineering TeamRead
- Flight Platform Engineering TeamRead
- Payload Integration TeamRead
- AI and Software TeamThis page
- Testing and Quality TeamRead
- Delivery and Training TeamRead
- After-sales and Customer Support TeamRead