Research — dissemination page

Towards auditable zero-shot detection.

The citable record of our Innovate UK project (Sovereign AI Pipeline Phase 2): what we are building, why and what we publish as the work matures. We publish the direction of the work here; the method detail itself stays commercially confidential.

Inspectable by design

Detection built from parts that are themselves inspectable concepts, so an explanation is a property of the architecture rather than a post-hoc account. EXPANDED ABSTRACT — FORTHCOMING

Zero-shot defect classes

Detection that extends to defect classes a production line has never labelled, with little or no retraining. EXPANDED ABSTRACT — FORTHCOMING

Evidence humans can read

Translating learned features into human-reviewable visual concepts. EXPANDED ABSTRACT — FORTHCOMING

The traceable decision path

Following a detection through the model to the evidence SparkSoft Assure renders beside every call. EXPANDED ABSTRACT — FORTHCOMING

Assure applies this work to image inspection today. The same explanation layer is built to extend to tabular, time-series and text data — towards a multimodal platform whose internal features, even those with no intuitive human name, are rendered as evidence a reviewer can assess. Read more on the Technology & Assurance page.

Publications, preprints and talks will be listed here as they appear.