Merantix: Empowering machine learning experts with Google Cloud managed services
About Merantix
Founded in 2016 by Adrian Locher and Rasmus Rothe, Merantix designs and delivers machine learning solutions for industries ranging from healthcare, to automotive, finance, and more. Based in Berlin, its highly qualified team of engineers works to bridge the gap between academia and practical challenges in the real world.
Tell us your challenge. We're here to help.
Contact usThe machine learning experts at Merantix leverage Google Cloud products and managed services as they find AI solutions for companies across all industries.
Google Cloud results
- Reduces the need for third-party code testing with GitHub integration on Cloud Build
- Saves an estimated 30% to 40% of engineering time with Cloud Bigtable as a managed service
- Accelerates an estimated 90% of basic startup tasks with managed services
- Cuts the cost of mistakes to encourage experimentation
Builds 3x faster with internal code testing
Founded in 2016 to bridge the gap between machine learning research and industry, Merantix has already established itself as one of Europe’s leading platforms for commercializing AI. Its highly qualified, Berlin-based team of engineers tackles all types of decision-making problems, innovating solutions for everything from automating breast cancer screening to improving quality control in factories.
“Merantix is an incubator for ideas,” says Nicole Büttner, Head of MX Labs at the company. “It’s a space where tech and industry experts come together to create real-life business value with state-of-the art machine learning, be it for industry clients or as a basis for products that can then spin off into independent companies.”
The team has set up two such companies so far, MX Automotive and Vara Healthcare, while its MX Labs operates as a consultancy, working with household names such as Volkswagen, Bosch, and TUV. The company explores new use cases and builds fresh applications of deep learning, such as spotting defects on a factory production line. Google Cloud plays a key role in the whole development process, including storing customer data, standardizing pre-processing for computer vision tasks, and training complex models in a short time.
“At the center of every one of our business cases, there is a simple data science challenge,” says John McSpedon, Machine Intelligence Engineer at Merantix. “When we look at self-driving cars, we can see a version of the classic task of distinguishing between a cat and a dog in a photograph. It’s virtually impossible to describe in words the small differences in the shape of an ear or a whisker that we pick up on as humans. A self-driving car needs to distinguish between a thousand different things in an image, between vehicles, signs, pedestrians, and understand how they’re moving. And there are whole new problems on top of that, such as recognizing whether someone is about to jaywalk.”
Training a computer to identify cats, dogs, cars, or wayward pedestrians means providing tens of thousands, even millions of examples for it to analyze. And workloads on that scale are standard practice for building machine learning models, as John explains: “These methods only work when you have very large amounts of data to work from and fast computers to read that data fifty times over. That means a lot of our budget is actually spent on computational power, using GPUs for days at a time. At the same time, there are lots of auxiliary challenges, too, such as importing data from an antiquated healthcare system or streaming terabytes of data from a self-driving car.”
“We work closely with the open-source world and open-source products such as Apache Beam and TensorFlow, which is the basis for our entire infrastructure. It’s clear that Google Cloud, with its focus on open-source technology, particularly TensorFlow, is the best platform to support our approach to machine learning.”
—John McSpedon, Machine Intelligence Engineer, MerantixMerantix needed infrastructure that would empower its engineers rather than act as a drag on development, even as the team took on diverse challenges. To do that, it looked for flexible, scalable, cost-effective infrastructure with managed services, ready to minimize the maintenance burden and enable experimentation.
“We work closely with the open-source world and open-source products such as Apache Beam and TensorFlow, which is the basis for our entire infrastructure,” says John. “It’s clear that Google Cloud, with its focus on open-source technology, particularly TensorFlow, is the best platform to support our approach to machine learning.”
Making the most of managed services
Every new project Merantix takes on brings different demands on infrastructure, from streaming data, to ingesting archives, or running large, temporary tasks. To do that, the company looks to benefit from accurate billing that gives it the flexibility to test new concepts, scale up and scale down without punishing costs.
“The granular billing of Google Cloud was a real draw,” says John. “My previous company used another cloud provider, and within two years its spending went from $2,000 to $1 million a month, and its database still had the same problems. I really didn't want to experience that nightmare. If you're testing a script and it has a small bug, the machine you run it on will spin up and shut down within the first minute. In 2016, other platforms would bill for a full hour in that situation, but Google Cloud billed by the minute, or even the second on some products. That makes experimentation easier, because mistakes aren’t expensive. It shows a sensitivity to what developers really need.”
Even for ambitious start-ups at the cutting-edge of AI, everyday tasks need to be addressed. “For every start-up, about 90% of the work is very similar,” says John. “You need to run a website, make sure you don't get hacked, provide logins for customers, and so on. With Google Cloud, we can do that 90% very quickly, because it has prepackaged solutions and managed services.”
More than any other managed service, Merantix uses Cloud Dataflow, which ingests data from a wide variety of sources and prepares it for machine learning models. “In a traditional framework, we would have to tell an engineer to do that job,” says John. “The cluster they would need to set up would have to run for at least a day, and all of our researchers would have to go through that engineer, creating a bottleneck. With Cloud Dataflow, we can request 100 or 1,000 machines for a processing job, and it will spin them up, charge you for the minutes the machines are used, and delete them as soon as the job is over.” That keeps costs down for clients and gives engineers more freedom to deal with challenges fit for their expertise.
“In my previous machine learning experience, we spent 30% to 40% of our engineering time maintaining a giant Apache HBase database,” says John. “That maintenance was expensive and caused significant downtime, so now we just pay for Cloud Bigtable: no maintenance, no downtime, a similar cost for storing data, and better performance, too.”
Testing code internally, without third parties
“Testing code with Cloud Build, we don’t need to create data samples for a third party or fiddle with passwords. Everything stays inside Google Cloud, so we have half as much code to write, the build is three times faster, and it’s cheaper. It’s a win on all fronts.”
—John McSpedon, Machine Intelligence Engineer, MerantixA key part of the team’s innovation process is that new code needs to be tested, and Merantix uses the GitHub integration on Cloud Build for continuous integration and deployment. “Before, we would pay for third parties to test our code, and would need to transfer data at a much slower rate,” says John. “If we had a lot of confidential, proprietary data, such as from a hospital, we needed to spend time choosing a small, anonymous dataset that we would be allowed to send. Testing code with Cloud Build, we don’t need to create data samples for a third party or fiddle with passwords. Everything stays inside Google Cloud, so we have half as much code to write, the build is three times faster, and it’s cheaper. It’s a win on all fronts.”
Made for machine learning
Merantix has used Google Cloud for all of its work so far, and was recently approved as one of the first customers in Europe to use the new Transfer Appliance to collect and upload medical data from radiological partners.
“As we explore new partnerships, our existing projects are set to spin off into new, independent companies over the next year, each building on Google Cloud at its base. Going forward, we’re especially excited about Cloud TPU, a custom chip just for machine learning, built right into the platform.”
—Clemens Viernickel, Entrepreneur in Residence, Merantix“We get data from a lot of hospitals,” says John. “Often they have hundreds of terabytes of medical files that we have to anonymize before they can be uploaded. We've mostly done that by going on site and installing servers alongside in-house computer systems, uploading files one-by-one to Google Cloud. Because the Data Transfer Appliance automatically integrates with Cloud Storage, that process is much more scalable.”
Meanwhile, the intellectually curious team at Merantix is expanding, on the lookout for fresh challenges where deep learning can make a difference. And from new technology to simplified services, Google Cloud will be at the core of any new project.
“As we explore new partnerships, our existing projects are set to spin off into new, independent companies over the next year, each building on Google Cloud at its base,” says Clemens. “Going forward, we’re especially excited about Cloud TPU, a custom chip just for machine learning, built right into the platform.”
Tell us your challenge. We're here to help.
Contact us
About Merantix
Founded in 2016 by Adrian Locher and Rasmus Rothe, Merantix designs and delivers machine learning solutions for industries ranging from healthcare, to automotive, finance, and more. Based in Berlin, its highly qualified team of engineers works to bridge the gap between academia and practical challenges in the real world.