01 February 2022
Another day in the model
For our Technical Lead Damon Maria, no day is the same: updating the AI models, developing and analysing data the systems deliver, studying every square centimetre of a hide to any discrepancies in system performance. Speaking the language of master leather graders as well as the language of code, he is a vital piece of the puzzle to our automation of wet blue grading.
A bit of a remote expert grader by now, Damon still learns new things about the tanning industry every week: ‘For an industry that goes back in history so far, it is surprising that all are after the same end product, yet every tannery does things in their own way. Our solution was built on scalability, but that unique character means we spend considerable time on the tannery-specific grading of each new customer. Each tannery attributes a different importance to the defects we can detect. Some may focus entirely on veins, whereas others really wish to zoom in on brand and various types of scarring. They then have their own way of categorising various grades and determining the benchmarks for when a hide fits any given category. We can spend several months training this model, which is custom developed for the customer.
By feeding the system large amounts of defect data we capture onsite, and labelling the hides according to the customer’s requirements, the AI starts to recognize the patterns. From there, it will start to mimic the decision-making of a ‘master grader’ – but with superhuman vision.’ Sometimes this precision causes some confusion, as a certain rule that the model was trained on (based on customer input), turns out to not be followed consistently by graders on the floor. In some cases, this causes a visible discrepancy in how grading results turn out for the customer pre- and post-system implementation. It’s why a close collaboration during roll-out is essential in finetuning a customer’s grading rules. It also means the tannery’s project team internally must be fully aligned in how they define accuracy: ‘If management wants to see an absolute rule-system, yet those on the floor have a softer approach to grading, this needs to be crystallised prior to us customising the model for them.’