How Walmart Labs Used AI to Improve Retail Item Coverage

Sharon L. Hadden

November 25, 2019

7 minute read

Walmart Labs recently published an article regarding their work with Samasource, where we delivered high-quality training data to further their machine learning initiatives. 

The article highlights how Samasource has covered more than 2.5 million items, improving Walmart's retail item coverage from 91 percent to 98 percent. 

In this interview, I chat with Taylor Rouleau, Project Manager at Samasource on how she and her team made sure Walmart's machine learning model was set up for success.

 

Transcript: "Ask Me Anything" Part 1: How Walmart Labs Used AI to Improve Retail Item Coverage

(00:04) Hello, and welcome to the Ask Me Anything series by Samasource, where we interview subject matter experts working in artificial intelligence. I'm your host, Sharon L. Hadden, an AI enthusiast and content marketing manager at Samasource.

Project Manager and Social Intrapreneur, Taylor Rouleau has a passion for social responsibility in business. At Samasource, Taylor leads production teams in five delivery centers across three countries. When asked, how did Walmart Labs use AI to improve their retail item coverage from 91 percent to 98 percent, here's what Taylor had to say.

(00:45) Hi Taylor and thanks for joining me today. I know in a recent Medium article, it mentioned how Samasource has improved Walmart's item coverage from 91% to 98%, and I'm just curious of what needed to happen from a project management perspective to achieve that. 

(01:05) We've been working with the Walmart team for the past few years and taking their retail item coverage from 91 up to 98% has been a really great milestone for us.

The main tasks that we do for Walmart is to tag items, items in their multimillion item catalog. So they'll give us about 6,000 product types and they update that about every month on average. We look at the items and say, okay, this is X product type. It might be a bike pump or a casual t-shirt.

Our data training agents learn and practice remembering a pretty large portion of that taxonomy so they can refer to it really quickly. One person who does really thoroughly know the taxonomy is our team lead in Nairobi, Johnas Wandera.

He's been really instrumental in training our team of 63 agents to tag everything quickly and accurately. Honestly, as project manager, I just make sure we have everything we need and clear and good communication with the client, to allow Johnas to do his excellent work with the team on the ground.

(02:09) You know, when you were saying, you know one person who's memorized 6,000 product types, I'm thinking Johnas must have some memory.

(02:21) He really does. But honestly he's been working with it for going on 10ish years now. Back from when we first started with them in 2009, so he's had quite a bit of time and practice to get very familiar with it.

(02:35) I would love to know was that 7% increase a targeted goal or just a sweet side effect? 

(02:44) Probably a little bit closer to the later. It really wasn't a stated goal, but the Walmart team has been a really great partner working together with us to deliver that 7% increase. It really makes a difference to have a dedicated platform for project setup, workflow and delivery, and the customized training we've done with our agents makes them something like experts on the whole catalog.

(03:06) It even sounds like collaboration really made all the difference.

(03:11) Yeah, definitely. We work really closely with the Walmart team to make sure their expectations were being met, and the excellent production team that we have as well as our great platform, SamaHub made a huge difference.

(03:11) Well, what was the most important thing you would say you needed to do to deliver those results?

(03:32) I think the most important task as a project manager is making sure we really clearly understand the client's expectations and needs and then making sure those are clearly communicated to the data training team on the ground so that they're able to be successful in their work. So I guess just being a good communicator and liaison is the key.

(03:52) What are some common challenges you face to deliver training data?

(03:58) I guess the most challenging thing that will sometimes happen is that we'll be given image annotation work and the images we end up with when we're doing production are busier or more complicated than what the client gave us for our initial testing and training.

This can also happen if a client changes instructions. I had a client that originally told us that they didn't want us to tag anything smaller than 20 by 20 pixels in images. So we trained around that and had the agents prepared for that. But once we were heading into production, they realized that wasn't really applicable and they wanted us to tag everything.

We didn't think it was a really big deal at first, but pretty quickly realized that it meant that instead of three annotations per image, we had closer to nine on average. So every image was taking three times as long as we planned for. Thankfully we were able to do a project change request with that client and get them on the same page about the change in complexity, which allowed it to be successful and everyone was happy in the end.

(05:03) Wow. Well that can be a big change if requirements move away from what was originally scoped.

(05:11) Yup, definitely. Which is why it's so important to have a close partnership and good communication with our clients to make sure we're identifying those changes as they happen and continuing to meet the client's needs and expectations.

(05:25) Well, thanks again for joining me today. I think I just have one more question for you and that's, what do you love most about working in AI?

(05:35) I think the AI world is a lot less scary than lay people tend to think.  There's so much growth and change happening in the field, so I'm constantly learning and I have the opportunity to work on new and innovative projects pretty frequently.

There's also a lot of hubbub and controversy around AI, so it's really a privilege to work for an organization that makes sure we're training data ethically and also has a social mission. Working at Sama allows me the expertise to advocate on behalf of both our impact and our data training best practices and those of our clients as well.

Contact Us

This interview is the first installment of a new audio blogging series titled, "Ask Me Anything," where Samasource interviews subject matter experts working in artificial intelligence.

Sharon L. Hadden

Sharon is the Content Marketing Manager at Samasource where she's responsible for telling the story behind the company's impact sourcing mission and human-powered training data solutions. Sharon holds a MS in Integrated Marketing Communications and is passionate about helping social enterprises transform abstract concepts into results-driven marketing.