AI Skills Base

SkillsBase was approached jointly by DexGreen and Openreach to enhance its product support further. With fibre boxes being installed at every home in the UK, even the smallest margins of error can be fatal to network health. Therefore, even a 1% improvement has high benefits – it was easy to provide smart solutions and prevent thousands of customers each year from being potentially disconnected from the internet due to an installation fault. We have countless examples of ‘on-the-job’ fixes and long-term improvements in installation quality.

By deploying various trials over 18 months, we identified that an AI solution would be necessary to cater to the demand of engineering installation requirements. It would allow us to respond to all requests. Throughout these trials, we received hundreds of pictures of work quality, identified a trend, and assembled a prototype quickly using off-the-shelf image analysis tools. This prototype prompted a high-profile trial with over 5000 engineers to reduce the fault rate. The trial also dictated how we could best design our mobile app to aid the capturing process and how to describe the feedback given by the AI. E.g. auto-take picture once the image is correctly framed, highlight critical faults more prominently, further training, better content and documentation, allowing engineers to mark a fault as ‘ignored’, which allowed us to retrain the models for increased accuracy.

To achieve the solution we have today, multiple technical and social problems had to be solved:

1.Computer vision – detects small areas of an image                                                         Particular objects have tiny components, some as small as a piece of thread, so it wasn’t easy to achieve high accuracy. We had to change how we scanned the objects and wrote our software to wrap onto existing AI tools. This enabled us to detect damage with a high level of accuracy, rule out false positives and provide engineers with much more excellent feedback. The standard AI tool struggled to catch areas smaller than 5% of the entire image,
the region we saw was >0.5% of an image.

2.Interpretation and delivery of A. I Analysis                                                                                  We learned early on that interpretation of AI results is crucial. Analysing and testing the model over time is essential to gauge how and why specific results are given. This helped us interpret it in a meaningful way for the end user. Delivery of such results has to be near perfect – clearly distinguishing between a fault, a warning or a nice-to-know problem. It can be the difference between a happy and an angry end-user. For E.g. we have found throughout our trials that mentioning ‘Did you know? You can do this…’ was often interpreted as a fault rather than a helpful tip. We were also able to add a new fault to our models and see its benefits the very same day, as engineers were shown this new fault and were able to react accordingly.

3.Social impact of wrong detections
Fear of AI technologies cannot be ignored, and we tried to do everything we could to make the end users feel safe when using Virtual Buddy. We provide their data back to them where possible and collect only necessary information – nothing more.

4.Fit Virtual into an engineer’s tool kit without disrupting jobs and with no upfront capital cost                                                                                                                                                           Virtual buddy needed to be simple, fast and effective to be adopted by engineers. Our UI, UX, and AI analysis had to be sleek and fast. We have also developed software to help support other products without any subject matter experts by using a smart form builder to help communicate the fault
detections effectively. Users can build an AI model by applying a simple logic pattern: Faults to detect, Minor issues to flag, the severity of detections, the probability threshold of each fault detected, if enhanced processing is to be used and many more. Our AI Models can be deployed very quickly, ensuring that it’s a practical and fit-for-purpose solution.

5.A limited supply of AI-trainable photos
A limitation for platforms of this kind is generally the number of photos required to start the image bank; we have overcome the technical hurdles, and our technology enables us to build a functional, field-ready AI tool with as few as 20 photos. We have incorporated smart image enhancements, enhanced image processing and image multiplication features to automatically develop and train an accurate image bank. The developer’s efforts are minimal, and no coding is required.