by Poojitha Jayadevan

Sheela Foam speeds customer service with machine learning

Feb 17, 20215 mins
Artificial IntelligenceMachine LearningMachine Vision

An Indian mattress manufacturer used machine learning to identify defective mattresses in photos sent by customers, at a time when staff couldn’t inspect them in person.

Charu Bhargava, VP-IT at Sheela Foam
Credit: Charu Bhargava

Just a month into the COVID-19 related lockdown, mattress manufacturer Sheela Foam saw a growing pile of unattended customer complaints. The pandemic had left the field force unable to visit dissatisfied customers to inspect their mattresses, and all a customer could do was register a complaint with the customer care team—and wait.

Sheela Foam’s customer care team, the Sleepwell Care Cell, identified the problem during one of their functional calls and, together with the sales team responsible for inspection visits in pre-COVID days, concluded the company needed a quick technology fix.

As it turned out, the IT team led by Sheela Foam’s vice president of IT, Charu Bhargava, already had experience with some of the key technologies needed to solve the problem. The company already used 3D image processing, image analysis, and machine learning to spot problems on the production line.

“The customized 3D cameras throw images of each mattress produced. These images are then compared with the images of ideal mattresses. The system checks if the mattresses pass all the quality parameters before they proceed to the next step of production,” says Bhargava.

To apply the same technology to inspecting mattresses in customers’ homes would require changes in the 3D object creation process to make use of images taken by customers using their mobile phones.

Pushed by the pandemic

While this could also have been done before the pandemic, there had been no reason to, says Bhargava.

“From the customer’s point of view, rolling it out before COVID-19 didn’t make much sense because customers are hesitant to make efforts when it comes to such pain points. They would rather have someone from the company come and check it at their residence instead,” she says. “Since COVID ruled that option out, and the fear of stepping out crept in, customers willingly adopted this solution.”

With customer complaints piling up Sheela Foam had a tight timeline for the project, 15–20 working days, with all the coding done in-house using Agile methodologies with some expertise from Staqo World, Sheela Foam’s technology partner and also a subsidiary of Sheela Group.

Around 12–13 team members worked on different aspects of the project, says Bhargava. “There is a three-four member group who worked only on image analytics, their expertise being the Python language. Two-three people who work with machine learning and image analytics together. Another team of 3 members who have expertise in React. Then there is a group of three who were involved in the 3D object creation and analysis process. There is also a separate dedicated group that is responsible for linking these technologies together through APIs and various other integration platforms. It was a mix of different technology skill sets.”

The team created an initial training database of over 40,000 images for its machine learning model, including some of mattresses identified as defective before they could be sent to dealers for sale. These were used to train the engine to detect problems such as sagging, bulging, and color fading.

“We started with a simple video call with the customers to look at the product. But then it started getting difficult because asking customers to share all the documents through WhatsApp or email wasn’t very easy. We needed a quick and easy solution,” says Bhargava.

From there, the company moved on to collecting basic details from customers through WhatsApp before sending them a link to a video tutorial explaining how to take and upload pictures of their mattress for inspection. The images are then sent to the machine learning engine for analysis.

Difference engine

“The engine is trained so that it can differentiate the defects,” says Bhargava. “For instance, it’s natural for mattresses to take the impression of a body during use, and doesn’t necessarily mean it’s an issue of sagging. The system is trained to differentiate between body impressions from an actual defect like a depression in the mattress.”

With the new image processing system, the size of the beds could also be measured accurately. In India, most beds are custom-made and do not come in standard sizes, so mattresses too have to be made in multiple sizes. The system developed by Bhargava’s team also includes a size measurement tool that helps in making fitted mattresses, resulting in zero alterations, savings in logistics, and improvements in quality.

Most importantly for dissatisfied customers, Sheela Foam reduced the time taken to redress complaints from 7–15 working days to three. “Earlier, if the mattress was defective, then we issued a voucher to the customer and they could get a new mattress from the nearest dealer. So it took over 15 days to solve the issue. But with this solution, within three days, we place their order online. And with our logistics partner, we deliver the customized mattress to the customer directly from the factory,” says Bhargava.

Sales staff are benefitting too as they are spending less time inspecting mattresses.

Convincing the managing director and the group CIO to deploy the project was easy, says Bhargava. “We involve the finance team to do the cost-benefit analysis only if the budgets are very high. This project didn’t involve any major budget allocation, and hence getting it approved wasn’t a problem.”