Amazon is obviously the leader in cloud and has been dominating the market for a long time. Using its learnings from its own internal IT systems, Amazon is now using machine learning to solve pragmatic problems. In an interview with Express Computer, Madhusudan Shekar, Head-Digital Innovation, Amazon Internet Services Private Limited, shares how his firm is now trying to democratize AI and ML technologies to the hands of every business.
Some edited excerpts:
How does Amazon use technology internally?
For over two decades, Amazon has been using artificial intelligence and machine learning to deliver better customer experiences. We are a very customer obsessed organization, which effectively translates into the fact that when you get data, we actually work backwards on how customers actually want it to be. And, then we actually build technology that can be applied and used proactively.
Through the years, we have built a lot of AI/ML technology capability. When it became more realistic to start using it on the cloud, we started offering capabilities for customers to do the same thing. The most major release that we did was in 2017 when we launched our machine learning platform called Amazon SageMaker which fundamentally enables every developer use Amazon SageMaker to build machine learning capabilities for their organization using any framework.
TensorFlow is the most popular framework in the market. We love TensorFlow. 85% of TensorFlow that runs on the cloud runs on AWS. You have complete flexibility over any framework and one of the important things that happens in AI and machine learning is the fact that you start off with one framework with TensorFlow, but then you may eventually find out that, you know PyTorch might be a better framework or some of the team may need to use PyTorch for certain app outcomes that they want to produce. So we optimize all the different frameworks to run on AWS and give complete flexibility to the developers. Remember, AI ML is still in the very, very early days, sort of tying yourself to one particular framework as an organization or an enterprise or a startup, etc. may not be the best thing to do. We have what we call AI services, which are a set of capabilities that allows an organization to take advantage of AI/ML, without knowing much about AI/ML. We are trying to democratize AI.
How has Amazon democratized ML?
With AWS, you really do not need any machine learning or data scientists in your team. We are literally bringing your data in as input. It goes through a sequence of operations, and then you get an outcome. And that outcome can be consumed by developers so that they can come to your mobile app or your website as recommendations for you. The same thing we do with forecasting, and forecasting is very important. Let’s take a really simple example of IPL matches. If somebody has to determine the amount of food delivery orders that are going to come, that will be a forecasting requirement. I will also give you an Amazon example – when we do Prime Day, one of the biggest challenges is that, we still have to do two day fulfilment on Prime, which means we have to forecast what customers will buy, how many people we will need to do deliveries, where the deliveries would occur and all this is- forecasting. The big engineering capability we have built, the machine learning capability we have built is now available for customers so that they can do their demand and supply chain forecasting and human resource forecasting. The same capability is now available as part of Amazon, personalized as a service to the customer with Amazon Polly.
Can you give us some examples?
I can talk about PolicyBazaar in India, who has been using Amazon Polly text to speech service. And Amazon Polly service actually, is very good because it does not only English, but also does English and Hindi. And in Hindi, it understands English text, which is Latin characters, and it understands Devanagari script. It also understands the script that they call Romanagiri. Romanagiri is basically Latin characters, but Hindi words, which is literally how a good part of India texts, right? So what we can do is you can type Romanagiri text and Polly will actually pronounce it in Hindi and you can actually mix them in a single sentence and Polly will identify the voice. It will actually understand some of that and will be able to pronounce it naturally in a one single sentence. That’s already developed and they are using it.
The usage is a very interesting one, for example say a request for a policy file and you were supposed to submit a set of documents and you missed a document or one of the documents wasn’t correct. Now, Amazon Polly will actually make a call to you and give you information saying that the current documents you provided are for the month of February, March and April and that Policybazaar actually needs documents for the month of May and June. The whole thing will actually come out as one natural speaking sentence. And, they do about 100,000 such calls using Amazon Polly.
Shaadi.com actually has a very interesting problem, right? People go in, put in their personal profiles and the challenge is when people claim that they are someone else or they may have the wrong information. That could have been deliberate or not deliberate, or if we have a situation where somebody says, ‘I am male’ but uploads a female picture or vice versa, or if you’re saying you are uploading a profile picture, but you are actually loading a group selfie. So, what used to happen previously was they had this validation work pipeline – they will see the picture that comes on Shaadi.com, somebody would look at the picture and say that we are good to go; human intervention was required. Now, substantial part of the workflow is being automated to an extent that allows them to very quickly validate what’s in the picture, whether it’s acceptable, and anything that additionally requires human validation goes to a human. If a photograph does not require too much work, everything is clean, there is one face, one person, face is clean, I can see two eyes, etc.. it goes through all the way to the website in just a minute. This is how Shaadi.com does it.
From an enterprise standpoint, can you give us some examples?
Broadly, if you look at AI/ ML, the number one thing that the businesses go after is our customer experience. This is because deploying AI/ ML services or capabilities to improve customer experience gives them a faster drawn investment. A lot more methods and best practices are available in that space for you to hack that problem. Whether it’s customer service, sales, propensity modeling, marketing, automation, etc. Next comes industry specific cases. For example in the case of manufacturing, it could be warranty analytics. In the case of FinTech payments, it could be payment fraud. And traditionally, payment for analytics, payment for processing was actually a rules based engine. Now, we are actually building deep neural networks that look at the graph of your payment. So one from the merchant side, what kind of transaction they have on their side and on consumer side what kind of transactions you have, and geographic propensity and building the neural and structurally identify and see what it is that’s second.
Third is insurance company for claims processing. So there are insurance companies allowing customers to take pictures of their car and send it in for their claims processing automatically. That, and insurance processing claims, automation, underwriting automation, etc. – these are the areas that we see in that space. All of them are doing customer experience in some shape or form. I have not come across somebody who’s saying “I’m not doing it” because everybody wants to learn the basic level AI. Services like Amazon Polly, Amazon Rekognition are focused on generating a specific output, in AI and ML. ISVs like Freshworks, Intuit, etc. then build models on these that can actually benefit other industries and marketplaces.