Real incidents. Real imagery. Measured results.
Two kinds of evidence: how Ralliyo's AI performs on publicly labelled track photographs, and what published accident investigations show about the inspection evidence that Ralliyo is designed to keep together.
How accurate is the AI on labelled defects?
150 labelled images from ZITMS (Roboflow Universe), scored against the labels.
What the investigations found
Track screws holding rails and baseplates to wooden bearers broke over weeks or months without being picked up by inspections. The rails spread apart under traffic until a wheel dropped between them.
RAIB 07/2020 · 6 May 2019Track twist from an embankment losing ballast supportWillesden High Level Junction, north-west LondonA degrading earth embankment and lost ballast support produced a significant track twist on a tight curve. Information about the earthwork and the track sat with separate teams.
RAIB 10/2015 · 2 October 2014A known rail defect left on a 52-week replacement windowPorthkerry, Vale of Glamorgan line, WalesA rail defect was found two months before the derailment, but the standards then in force did not treat it as urgent, so replacement was scheduled within 52 weeks. The rail failed first.
Contains public sector information published by the Rail Accident Investigation Branch and licensed under the Open Government Licence v3.0. Licence. These walkthroughs are reconstructions for illustration and do not claim Ralliyo would have prevented any incident.