Shipper is one of the fastest-growing tech companies in Indonesia, working to digitize Indonesian logistics and enable cost-efficiencies at scale nationwide. Since its 2017 founding, Shipper has built a vast network of fulfillment centers and partnered with hundreds of local delivery companies across the country in pursuit of this goal.

Trax Retail offers advanced solutions for dynamic merchandising, in-store execution, shopper engagement, market measurement, analytics, and shelf monitoring to help drive positive shopper experiences and unlock revenue opportunities at all points of sale. As a global pioneer serving customers in more than 90 countries, Trax Retail leads the industry in innovation and excellence through the development of advanced technologies and autonomous data collection methods.

Trax Retail’s solutions are based on AI models trained using neural networks to recognize products on the shelf in supermarkets and grocery stores. The company’s AI-Engineering team supports these models, focusing on researching infrastructure, training the AI, and retraining the models as needed to maintain accuracy.

They also manage the production pipeline to monitor for issues.

According to Tzoof Hemed, AI-Engineering Team Leader at Trax Retail, this entire deep learning process is typically very tedious. It requires collecting and processing data, training the models, monitoring that work, and then comparing it with current solutions. For instance, it could take weeks for a data scientist to complete.

This type of process is particularly challenging at scale.

Each Trax Retail customer requires a unique AI model, and that adds up quickly. As the company grew, conducting large scale training across the entire customer base required team members to write down the IP addresses of the servers they used so they could monitor each of their experiments and runs to see the outputs of that training.

And the process doesn’t stop there. The type of deep learning pipelines the Trax Retail team builds are quite long. That means they’re composed of multiple tasks, and a failure in a single task can affect all the others that follow.

In turn, this situation makes debugging complicated without a clear view of the specific scope of each task and its larger impact.

However, Hemed shares that not only did his team not have this detailed view, but their network was actually more of a “black box,” which made the debugging process even more difficult.

Recognizing these challenges, Hemed and his team began the search for a solution that would allow them to increase the scalability of their work to support Trax Retail’s growing business.

“We were looking for something that would orchestrate all the different steps that we take, which could be dozens or even hundreds depending on the implementation. We wanted to save our team time so each person didn’t have to monitor their servers or tasks at the individual level. We wanted a single platform where someone could just log in, see all their experiments, confirm all their runs went through, and easily search for details like the outputs of different tasks and the results of training,” Hemed explains.

The search for a solution to automate the deep learning process, including collecting and processing data, training data, monitoring AI models, and debugging, led the Trax Retail team to Databand.

Databand’s proactive data observability platform offered exactly what the Trax Retail team needed. Specifically, it helped solve two key challenges Hemed and his team faced:

1. Automating the data pipeline, including optimizing how data gets collected, processed, and trained.
2. Monitoring data pipelines to understand accuracy and support debugging needs.

Based on these capabilities, Hemed shares that the decision to implement Databand was an easy one: “Databand gives us a single platform where everyone can talk about the same outputs, and that’s a key point for us. It allows us to see all of our deep learning training pipelines in one view to troubleshoot any issues, plus we can easily compare pipelines to see what’s working and what’s stopped working.”

The Trax Retail team implemented Databand alongside the company’s Kubernetes engine, which allows for true automation. Hemed reports that Databand now implements all requests in Kubernetes and ensures that runs are complete, which means his team doesn’t have to make any deployments in Kubernetes.

“We just run a command line. It’s that easy,” he says.

Trax Retail has seen incredible business impact since implementing Databand, most notably around reduced pipeline incidents and increased scalability of the AI-Engineering team.

Prior to using Databand, Hemed reports that about 60% of his team’s pipelines experienced data incidents. Since implementing Databand, that number has dropped down to less than 1%.

He attributes this in large part to Databand’s ability to surface those incidents quickly and allow for immediate action. “Without Databand, finding those incidents would take days of debugging, including looking into all of the outputs and logging them. With Databand, we can find the incidents in minutes.

Having everything in the same place just makes it so much easier,” Hemed says.

In general, Hemed shares that his team is impressed with how many tests Databand enables them to run simultaneously and the downstream effect that has had on reducing pipeline incidents.

Databand has enabled the AI-Engineering teams to train and manage more deep learning models in a sustainable way. In turn, this means the team can handle more models for more customers without having to add more data scientists to the mix.

“One of the main advantages of introducing Databand was the scale for our team. Databand allowed us to jump from one data scientist handling a few models to being able to manage hundreds of different runs. That was a big leap for us,” Hemed explains.

Notably, this increased scalability has enabled Trax Retail to triple their customer base while keeping engineering costs flat since they’ve been able to do more with the same team in place. Hemed concludes: “In order for our business to grow, we need to be able to provide more models to new customers. If our team can’t train those models, that’s a huge bottleneck. Databand has alleviated those concerns for us. And our customers can notice the impact too: We can get them set up with a new solution faster since it takes less time and fewer people working on it to complete.”