MC Interview | AI to revolutionise e-commerce immersiveness: Meesho Co-founder and CTO

(Pictured: Sanjeev Barnwal, Co-founder and CTO, Meesho)

Indian e-commerce is a rapidly evolving market. Within its cutthroat waters, a tiny drop of blood is enough to spur brands into a feeding frenzy, largely dominated by big players like Amazon or Flipkart.

That doesn't mean the other players don't innovate. Take Meesho, for instance. The Indian e-commerce platform was started in 2015 by IIT Delhi graduates, Vidit Aatrey and Sanjeev Barnwal.

"We've built pretty solid tech, which has scaled to, you know, 140 million plus monthly active users. We've seen scale to the extent of more than four lakh requests per second," said Barnwal, co-founder and CTO.

In June, this year, Meesho's app crossed 500 million downloads, making it one of the fastest-growing shopping apps. But, being an e-commerce platform alone will not get you that much traction. After all, why would anyone choose Meesho over Amazon or Flipkart? This is where AI comes into the picture.

"Very few companies in India operate on the scale we do," says Barnwal. "We've made significant progress in leveraging AI across the entire stack for making more efficient decisions and creating real business value".

Not only has AI helped Meesho create more value, but it also helps them track counterfeiters and protect customers. The team first established a process in 2018, which involved a lot of manual handling.

"So it used to be rule engine based, which is simpler text matching and then manual checks as well," added Barnwal. "We formed our AI team in 2019, and this was the beginning of the AI journey at Meesho".

In a recently published report by the company, Meesho said that it removed nearly 42 lakh counterfeit products from its platform in six months. The technology also detected and blocked over 12,000 bad actor accounts, which have been blacklisted.

"If you think about it, the very basic form of text you can put in a search field is the name of a product," says Barnwal. "Let's say you search for 'Nike' and sellers who have tagged their products with that word will be shown in the results. But what if someone sold a product called 'Nyke' but tagged it with the original brand name?".

This is where AI image detection steps in. It can detect logos and check for known modifications of the same spelling, allowing Meesho to flag a counterfeit product.

"We maintain a database of known product logos from various brands, and match them with all incoming products, allowing us to decide whether we should list the product or not," adds Barnwal.

This allows Meesho to grow their accuracy rates and lower recall rates.

"Accuracy in very simple terms can be defined in terms of what percentage of my predictions are accurate?" says Barnwal.

"And recall is what percentage of fake products is our AI system able to catch?, AI algorithms and iterations help improve both these dimensions, and creates a much better experience for the user".

Of course, it also helps the team with giving correct feedback to its sellers and suppliers.

"For example - If your accuracy is bad, you'll end up blocking a lot of products, and this takes away any potential growth for the seller," adds Barnwal.

With so many AI models out in the wild, how does one go about building their own?

"These models are completely in-house," says Barnwal. "Yes, we leverage open-source stacks a lot, but all of our models are trained in-house and are deep learning based models".

The process of training these models are often iterative updates. For example - if there is an image-matching algorithm, the team continues to improve the code and using that to find new parameters to train the models on.

"We have our own multi-modal AI system," says Barnwal. "This means it can understand both images and text, and use them as input".

Naturally, training these models is a massive effort, and requires a trove of specialized data. Most of the training happens with public data gathered from the web.

"We scrape a lot of data from the web, legally available of course, and train our models," added Barnwal. "But being an e-commerce platform, we also have tons of data in-house. A lot of it is manually catalogued data, checked for relevancy and updated. It feeds into our algorithm and makes the model understanding richer".

Using the AI system, it takes the team an average of 72 hours to find counterfeit products and remove them. Barnwal says that the team is making steady progress to lower the detection time further.

As for ethics and principles guarding AI, Barnwal agrees there needs to thorough regulation and a standard for the technology's use.

"Creating regulations for AI-based systems is hard. You have to think about ethics, society, and technical problems," says Barnwal. "It's a multi-faceted problem. Take privacy for instance, AI models need data to run on but you need to make sure all of it is anonymized and has proper consent. Bias and fairness is another challenge to overcome, because AI systems tend to be biased towards the data you feed them".

Similarly, if AI is used in the hiring process, the system might favor certain resumes over others, based on background.

"Another problem is accountability," added Barnwal. "If an AI system causes harm, who will be responsible? Is it the developers? or the users who probably used it in an unintended way? Where does the accountability lie?".

Barnwal says that AI systems will give positive results 99.9% of the time but how do you handle the 1% failure cases? He believes regulation is critical for these issues.

Coming back to e-commerce, Barwal believes AI will play a huge part in disrupting e-commerce.

"Generative AI has huge potential to disrupt e-commerce. The industry itself has stayed largely the same, especially when it comes to discovery," says Barnwal.

AI can help make e-commerce more immersive. Barnwal believes that "some sort of hybrid between conversational and browsing based algorithms will be the end state for e-commerce".