Interview with Disney+ Hotstar & Myntra: Scaling AI from PoC to Production
Artificial Intelligence (AI) is a hot topic nowadays.
Some people believe AI enhances our everyday work and lifestyle—allowing us to devote our attention to more high-functioning tasks. Others believe it is a threat to our jobs and career potential.
Whichever side of the debate you’re on, it’s no question that AI, when driven by deep neural networks, can automate (oftentimes tedious) processes that otherwise require humans. In fact, tasks that occupy 45% of employee time could be automated by implementing demonstrated technology, according to McKinsey & Company research.
However, the ability to manage complex inputs is only the first step in this journey. AI-driven applications result in significant complexity across business systems.
Our HackerRank Host, Aadil Bandukwala, spoke with a panel of experts about scaling AI from proof of concept all the way to production. The panel included:
- Akash Saxena, SVP, Head of Technology at Disney+ Hotstar
- Ravindra Babu Tallamraju, Vice President and Head, Data Science at Myntra Jabong
- Harishankaran K, Co-Founder & CTO at HackerRank
Watch the full video below, or keep reading for some highlights from the discussion.
What does it really take to make a great application that's powered by AI?
First, you have to ask: What does it do for the customer? What does it take to build any great application?
AI is just another tool in my toolkit. It is a more enhanced algorithm—whether I'm using machine learning or neural nets to make things happen. With the onslaught of the data, there are certain problems that can only be solved by using AI techniques. To me, what makes an application powerful depends on the answers to the following questions: What does it do for the customer? Does it make the platform inclusive? Does it make the experience richer?
You must start by identifying that and then work your way backward. Indeed, for a lot of my use cases, AI is pretty dominant.
Ravindra Babu Tallamraju:
Anything that is manually done by a domain expert that needs to scale up is only scalable to a limit.
For example, let’s talk about some style grading. At Myntra, we take a particular garment, and have a fashion expert look, and grade each garment. This expert comes with a great amount of domain expertise—10 years of experience in evaluating each garment. Let’s say this style is awesome and we buy some n number of items of this garment. Because of the large volume of garments, we replace the expert with a probabilistic model, which performs as well as a human.
That’s an example of a great application. It’s scalable, and it should perform as well as the expert. It definitely can't exceed the domain expertise of a human, but it can perform just as well.
AI is expanding the horizon. The goal of AI is how well it can do or how much closer it can come to a human performing the same functions. That gap is reducing, day by day.
The most important thing you need is data. If you don't have enough data to make those predictions, then there is nothing beyond that. But now there are so many ways to collect data, so many places you can get the data and make those predictions and models. So, I think it requires some level of imagination.
There are some things that I always believed were not possible. Yet, there are new applications that emerge every day, and prove me wrong.
A recent creation is the CPT-3, which can predict what you're going to write. I have written mail using GPT-3. So that's one of the new enhancements that prove you have to unlearn everything and assume that everything is possible and go with that mindset.
How do organizations really determine if they have an applicable AI use case?
I want to share an anecdote that a very senior researcher shared once with me. You could have the best recommendation algorithms, but say your UX is broken, and suddenly all these recommendations are in a place where the customer doesn't see them. It's not the fault of AI, or it's not a failure of what the application was doing. This illustrates that all products should be built holistically.
Some great cases where AI is used is anything that you can do to make your product more scalable, repeatable, and consistent. Ravindra talked about fashion grading as an example. I may be a sharper fashion grader versus say, Ravindra. Ravindra may have other skills, but you can teach a machine some of these attributes. It's hard to get to where a human needs to get to, and it's very complex. You can add a whole bunch of layers and take 200 days to compute an answer that someone could give you in maybe a day.
Specifically, I am really blown away by self-driving cars. I'm a technologist, but I'm constantly amazed at what machines can now do and how aware they are. So, it's pretty fantastic. I think Harishan followed a similar theme in terms of what we pick up. We've tried to find a lot of operational efficiencies. When content comes to us, we process hours and hours of content every day. All of that content has to be looked at. Are there any anomalies? Is there anything that is objectionable legally? This is usually human-intensive. And our focus has been to find where there is a variance in how a human may perceive it.
There's always this fear that AI is going to eat our jobs. In reality, I'd say that it unleashes us to focus on more high-quality problems, versus trying to solve low-quality problems with just raw manpower. And we've used AI to great effect for all of our recommendation engines.
But we've also started to look at AI from traffic prediction models. We are using AI to put models in the client to figure out how we can do AVR, the bitrate calculation. So, when do you make a decision to switch from 360p to 480p? These decisions are made in the moment. The algorithm today says, well, can you do it now, like in this very instant? Are my last few samples good enough? Not realizing that maybe you are in a network area that's really bad. You may have a momentary peak, buffer, and then again come down. To my customer, that's a poor experience. I'd much rather not ladder up. That's an area of using it.
What are examples of successful AI use cases that continue to amaze you today and what are exciting problems you're solving on?
Ravindra Babu Tallamraju:
We have models that impact every customer touchpoint—from the moment they log into the homepage, to when the shipment has to be delivered, and on the return. Our challenge lies in ensuring that the returned item is exactly the same and is in the proper order.
Here’s another case: in our warehouse, we try to make the optimal assignment of shipments to Pickers. Here, the amount of data that is necessary is limited, and you could possibly approach it through a queueing theory model. The data is predominantly used to validate some of that. For the recommendations, as Akash is mentioning, you need massive data sets. You need to understand what the customer likes, what their previous purchases have been, and then can we recommend the right product to them. And this is where the massive data sets come and you know all the complexity of building a model.
As a business building a product, you constantly have to wow your customer. You have to incorporate the “customer delight” aspect. And AI is one strong tool that helps you do that multiple times.
Back when I had Yahoo Mail, I was super excited about how Gmail spam always goes to the spam folder. No spam would ever come to my inbox. Now, it's not impressive anymore. We have gotten used to it.
Today, as I start typing, it completes my statements. This is the next world. In a few years, I'll just type the subject and then it will fill the whole mail for me, and it's not going to be the great stuff anymore.
It comes to a point where if any recommendation comes in that I don’t like, say from Hotstar and Myntra, and if I don't like it, I might think there’s something wrong with me and maybe this must be the thing I need. You tend to trust in it so much that you go ahead believing this should be right and maybe I'm not in the right state of mind. So I believe that's the point where the successful case of AI, like self-driving cars, is one which would be the best successful AI use case.