In this multi-part series we're going to explore a real-life example of operationalizing your analytics and then quickly dive into the technical details to make it happen. The phrase Operationalize your Analytics itself encompasses a framework for realizing timely and relevant business value from models, business rules, and algorithms. It also refers to the development-to-deployment life cycle commonly used in the ModelOps space. This series is geared towards walking you through a piece of the puzzle, exposing your analytics as REST APIs, but to get a better feel for the whole picture I would encourage taking a look at these resources:
Fair warning: I get detailed here. This is a journey. I'm not trying to pitch the story for the next Netflix original, but I want to outline business challenges companies face and how they can utilize technology to solve these problems. The more background you have, the more the solution resonates. I also want to have a little fun and give props to my champion salesman.
Framing the use case
Our goal in this first part of the series is to better understand the value of building custom applications on top of SAS Viya. To illustrate my point, I'm going to use Tara's Auto Mall, a used car buying and selling business I hypothetically own. We incorporated the business back in 1980 when neon and hairspray were king and MTV played music videos. Kurt, pictured at our lot, was my best salesman.
This man knew how to spot a great buy. Every auction we'd send him to looking for cars, he'd find the gems that just needed a little polishing, a little clean up, and knew exactly how much to pay for them. Sure, we still bust on him today for a couple of heaps he bought that we never made a dime off, but for the most part he knew how not to get caught up in the heat of the auction. It's an easy thing to do while just trying to outbid the others that you end up paying too much for a car. After all, we've got to turn a profit to keep the lights on and we do that by bringing used cars from auction to our lot for resale. Over the years, Kurt learned which makes and models we could get the best profit margins out of, while also learning how to spot cars that might end up racking up too many unexpected costs. "Salt and sun damage will take a huge bite out of your commission" he'd always tell the new guys.
But these days Kurt, and most of the other great salesman I've had over the years, spends his days fishing in the middle of nowhere, happily retired. Sure, I've got a few folks still on the team that really know a good buy when they see it, but more and more often we're losing money on deals. As hard as these new kids work, they just don't have the years of experience to know what they're getting into at auction.
Sometimes a car will come up that looks like it's in great shape. Clean and well-kept, low miles even, and everyone starts spouting off bids, one right after the other. That'll make just about anyone think "they must know something I don't know, I've got to win this one!" But unless you really know what to look for, like Kurt did, that good-looking deal might end up being a dud. And while Kurt had experience under his belt to help his eye for that, keeping track of the good buys and the bad ones, learning from all those lessons - that's the best yard stick for comparisons when you're trying to make a quick call at auction. How much have we been able to sell cars with similar characteristics for before? Have they ended up costing us more money to fix up than they're worth?
Needing to modernizing our business
Okay, you get it. It's a challenge many are actively facing: a retiring workforce taking their years of experience with them, combined with agile competitors and market disrupters. Not only to better compete but also to cut down on poor "gut" decisions, Tara's Auto Mall decides to build an app. This way all their employees can use the app at auction to help make the best possible decisions for which cars to bid on and how much is too much before overpaying. Over the years they've bought a lot of cars, thousands, and no one can really keep track of all that data in their head to know how to truly evaluate a good buy.
The buying team should be able to pull up this custom app on their phone, put in the details for the car they're thinking about making a bid on, and almost instantly get back a suggestion on whether or not they should bid at all and if so, a maximum bid amount and some other helpful information. Those results should be based on both the historical data from previous sales and from business logic. And they should be able to record how much each car up at auction went for and it's characteristics, whether they won it or not, so we can feed that back into our database to make the models even more accurate over time.
How can Tara's Auto Mall improve?
Now that we've got the main requirements for the application that Tara's Auto Mall needs to build so the buying team can make data-driven decisions at auction, we can start to hash out how those recommendations will be generated. The solution they decided on is to use SAS Viya to create a decision flow to evaluate used cars up for bidding at auction by using that specific car's details as inputs and first checking them against some business rules that they've come to trust over time. This adds a layer of intelligence into the decision-making process that models alone cannot provide.
We heard a bit about the lessons learned over time, like looking for salt and sun damage, so we can layer this into the recommendation logic by checking for certain states where that type of damage is typical. This allows us to add more "human knowledge" into the computer-driven process and automatically alert the buyer that they should be aware of this potential damage. That decision flow then uses both supervised and unsupervised models to evaluate the car's unique characteristics against previous sales, the profits from them, and market data, to generate a maximum bid suggestion.
Here's the fun part
During an auction we don't have time to wait for some batch process to give recommendations and we don't want our buyers to have to log in and enter values into a written program either. Instead, we can publish the entire decision flow as an API endpoint to the SAS Micro Analytic Service (MAS), a “compile-once, execute-many-times” service in SAS Viya. Then we can have our custom application make REST calls to that endpoint with the specific car details as inputs, where it will execute against a packaged version of that entire flow we outlined. The result is a recommendation from both the business rules and models.
The SAS decisioning process
The app is part of the SAS decisioning process, and evolves over time. As more sales are added to the database through the app, the solution automatically retrains the recommendation model. It yields more-accurate suggestions as the data changes and then republishes to MAS so our custom app can automatically take advantage of the updated model.
I want to pause here and note that, while this example is for an application making REST calls to MAS, that is really the tip of the iceberg. One of the greatest advantages to the modernization of the SAS platform with SAS Viya is the sheer number of underlying services that are now exposed as API endpoints for you to incorporate into your larger IT landscape and business processes. Check out developer.sas.com to take a look.
Tara's Auto Mall has now decided on the custom application to help their buyers at auction and the solution to building repeatable analytics and using that to provide near-real-time recommendations. Next, we need to work through integrating the custom app into their SAS Viya environment so they can embed the recommendations.
Questions from the author
While you might not have gotten to work with Kurt, Tara's Auto Mall's best and most fashionably-dressed salesman, is your business experiencing similar challenges? Do you have homegrown applications that could benefit from injecting repeatable analytics and intelligent decision making into them?
Building custom apps on top of SAS Viya, Part One: Where's the value? was published on SAS Users.