India isn’t your typical e-commerce market. With the country’s high level of diversity in geography, languages, culture, literacy, and income levels, one-size-fits-all solutions won’t work. A service that appeals to people in one region may fail to attract them in another.
To deal with this complexity and scale our e-commerce approach at Flipkart, we’ve incorporated AI into our operations.
We’ve invested heavily in voice and vernacular technology, as I explained recently in my Tech Talk on AI Exchange. India is one of the most culturally diverse countries in the world. It has 22 scheduled, or official, languages. Today, our platform is available in English and 11 other languages, and offering so many options is an important part of serving customers in their own language. We translated roughly 5.4 million words for each of the interfaces by applying a judicious mix of translation and transliteration to offer a native, colloquial experience to customers
We’ve also poured investment into a tech-enabled supply chain, which has been pivotal in building customer trust in our e-commerce offerings. We’re doubling down on our efforts to bring a natural language experience and create shared value for millions of our customers, sellers, and ecosystem partners across the country.
Today, we have 350 million registered users in a market predicted to reach $200 billion by 2026, but we’re intensely focused on bringing in the next 200 million consumers.
While Flipkart faces some complexities that are exclusive to the market in India, we’ve learned several lessons that will be valuable to any e-commerce business that wants to integrate AI into its operations. Here are some of our key takeaways.
One advantage of using lightweight models is that they can run on a CPU rather than a GPU. Many machine learning models require a GPU, which adds expense to running the model. With a lightweight model, you can run your model faster and on less expensive hardware.
However, lightweight models can’t answer more sophisticated queries, so you’ll probably need a second layer, with a model that needs more powerful hardware, to answer those queries.
Having a good MLOps infrastructure not only saves time, but also allows you to experiment with different models. The time from finding a problem to creating and deploying a fix must be short, but it may also require building multiple models before you can find the correct fix. The longer it takes to perform those experiments, the greater the delay in deploying the new model.
Labeling is an activity usually performed by humans, and all humans are fallible. When you use the right tools, you have the quality controls needed to ensure that your model is getting high-quality data, because bad labeling means bad data, and bad data means bad models, and bad models mean a bad customer experience.
Active learning can help focus troubleshooters on specific problem areas in a model. If you have $10,000 to fix a model, do you want to spend it on areas that have problems, or do you want to label the whole thing all over again? Active learning allows you to systematically narrow down where to focus resources.
In addition, the time and cost of labeling images by humans can be immense. As an alternative, companies can use synthetic data, which includes computer-generated image data that models the real world. Visual effects technologies can work with generative neural networks to create large amounts of photo-realistic image data and to label that data automatically. This produces training data at a fraction of the cost and time required when the process is left solely to humans.
Analytics are essential for rapid troubleshooting of AI models. They tell you what’s going on in your model and help to pinpoint trouble spots when things are going badly. You can do troubleshooting without analytics, but that slows down the process.
Customers can ask a lot of questions, and those can jump from customer service to product discovery to decision making, and so on. You don’t want to have to hand them off to a different bot every time they change gears. By unifying the assistance experience through AI, customers can seamlessly interact with the e-commerce platform.
For example, if a customer asks a question that bridges discovery and decision, the system is smart enough to recognize that and provide a rich response.
Let’s say a customer looks at a product page for a phone and asks, “Does this phone have two SIMs?” The simple answer is no. The richer answer would be, “No, but here are three phones that do have two SIMs.”
Typically, customer response systems are designed around templates or rules. For example, a customer might ask, “How much memory does an iPhone have?” That might be addressed in a template that looks like “The _______ has _______ of memory,” where the system fills in the first blank with the device the customer is interested in and the second blank with the amount of memory supported by the device.
The problem with this method is that it doesn’t scale very well as complexity increases. Maintaining rules for thousands of items, to answer all of the questions that might be posed by a customer about a product, becomes incredibly complex.
A generative model can replace the need for all those rules by learning from customer behavior and other website data, and with an AI model it will continuously keep getting better and churn out the answers you want.
AI and ML can be strategically important to a successful e-commerce enterprise, especially as business begins to scale. By making the customer’s journey a seamless one, you can increase both sales and customer satisfaction. To reap those benefits, however, e-commerce companies must invest in the technology needed to bring those benefits to customers and sellers in their supply chains.