In these episodes of GAINS On, join GAINS Co-founder Bill Benton as he describes lead time sensing and prediction, empowered by machine learning through GAINS. We know there’s a lot of noise around AI and ML, everything from chatbots to social media ads, but what we focus on here is pragmatic ML and AI that can be operationalized to be resilient across teams and time, and has built-in bidirectional plumbing. So we’re excited to bring this.
(00:05): Hi, my name’s Bill Benton. I’m the co-founder here at GAINSystems. I’m excited to talk to you today in this GAINS ON segment regarding lead time sensing and prediction, empowered by machine learning through GAINS.
(00:40): Data’s noisy, right? So if you look at your supply history, there’s gonna be returns, there’s going to be negative amounts, there’s going to be dates that weren’t accurate. How do you work through that? We use some anomaly detection. So when we talk about machine learning for prediction, we also use it beforehand for detecting those anomalies and the data coming in, so that we do an initial cleansing. Secondly, we work tightly with practitioners that are trying to employ this with our GAINS Labs team to refine what we call the feature data. So this is input data, and we’ll usually iterate that two to four times to get to the point where we’re convinced that there’s vastly more improvement than degradation.
(01:46) So, traditionally, that’s been hindcasting. So when you look at your supplier performance, on time performance, split lines versus whole lines, etc., you might be punishing somebody who’s been improving rapidly and is likely to do well in the future. Or conversely, you might be rewarding somebody who’s very likely to have an imminent decline in performance. As you think about anything where you’re trying to negotiate and you’re looking at what they’re going to deliver, it’s much better to know what they’re likely to deliver next versus what they’ve delivered last.
(02:29): One of the achilles heels of machine learning models and making sure that they’re both operationalized and resilient across time and teams is refreshing. One of the reasons that we’ve built this into GAINS as opposed to simply importing other people’s models, is to ensure that. So we collaborate with customers on the feature data, and as you learn more about AI and machine learning, what you’ll learn is that the methodologies are less important than the data you feed it. And so what do is we ensure that there’s a consistent refresh of feature data and looking at new types of feature data as conditions change, but then also that the model automatically refreshes and retrains with available data and an available target to put that output as opposed to requiring, say, a decision scientist to rerun; it’s built to last.
(03:28): These would be environments where A: lead time is longer and more variable. If you are having to deliver parts or goods in far less time than you get advanced notice and you have a long lead time gap, this could be incredibly helpful. The more variable that lead time or lead time gap is across time, the greater the need for this. So if it’s steady, if it’s long, but steady lead time sensing and prediction isn’t gonna be as valuable. If it’s long and variable and variable cyclically this can be hugely valuable, as we mentioned to improve myriad operating performance measures as well as productivity for the planning team. Thank you for your time today. We know there’s a lot of noise around AI and ML, but what we focus on here is pragmatic ML and AI that can be operationalized to be resilient across teams and time, and has built in bidirectional plumbing. So we’re excited to bring this. We think it’s highly useful and easily implemented. Thank you.