Master student Patrick Michelberger shares his experience at BMW Group IT in developing and operating a data-based demand-forecast model for the mobility service ReachNow.
Since AI is becoming an increasingly important factor in software applications, I wanted to dive deeply into that topic for my master thesis, after having worked as Full Stack Developer for web and mobile applications in companies such as SinnerSchrader, ABOUT YOU and Roche. A friend brought BMW Group IT to my attention. Unlike the stale air you might expect from corporate environments, I was astonished to see how „IT-like“ and modern work-life is around here. I am part of a young and dynamic team consisting of Data Scientists, Data Engineers, Software Developers and Project Managers. Even though the team is embedded in a large organization, we enjoy lots of independence and the opportunity to work on exciting topics. My teammates have a broad spectrum of academic backgrounds spanning physics, math, statistics and computer science. The interdisciplinary set-up is actually very helpful, because the team does have hard nuts to crack – often starting from scratch to pave the way for new mobility experiences. The Data Science department’s mission is to develop data-driven applications based on machine learning to improve established processes along the entire value chain of the BMW Group, e.g. product development, production, maintenance, mobility services or after sales.
Data Science as Software
For the master thesis, I am engaging in a project for the BMW Group’s ReachNow, a mobility service with both a car-sharing and ride-hailing service. Our first target was to optimize ReachNow’s ride-hailing pilot in Seattle. Particularly, the scheduling of drivers in an area where demand prediction based on machine learning is key to success and therefore a highly appreciated contribution to the business. We started in a true pioneer mode, envisioning what ReachNow needed and how the model could be set-up. At the same time, we needed to create something that is easy to implement and modify. The first project phase was actually quite hacky, because we had to get people from different departments and functions such as R&D and operations “on the same page”. In addition, the live data integration wasn’t exactly a piece of cake, but we got it worked out and managed to show something tangible early in the process. After we received confirmation from our stakeholders, we continued building on top of that.
The tool itself is a big matrix, showing the different Seattle neighbourhoods on one axis and the respective time slots on the other. This means for all parts of the city and at every hour, the tool indicates higher or lower vehicle demand based on the underlying prediction model which take factors such as weather, events, points of interest or competing taxi stands into account. Furthermore, it is possible to drill down into every cell to see the detailed metrics. That is necessary to deepen trust and understanding regarding the machine learning model.
You build it, you run it
The challenges did not end once we had developed the tool. Deploying it is actually quite as tricky, because the data scientist needs to retain the option to make adjustments in case of performance problems. This feedback loop is essential for improving the quality of the model and guaranteeing its long-term success.
In the accompanying blog, you find more information on how you can leverage cloud services such as AWS and connect the data scientists with the day-to-day operation of their models. You will find a step-by-step approach how to deploy a pre-trained scikit-learn model in combination with AWS Lambda and API Gateway. It is an easy, secure and low-cost solution that is based on a serverless architecture. Serverless is a hot topic today. A nice introduction by Martin Fowler can be found here. In the project we use AWS Lambda as a ‘’Functions as a Service” (FaaS) provider to load the trained model from S3, make the actual prediction, send back the inference result to the client or optionally save it to S3 or other AWS services. For deploying the application with a single command, we use the serverless python framework Zappa. It handles all the necessary configuration and deployment automatically: API access management, security policy generation, precompiled C-extension, auto keep-warms, oversized Lambda packages and many more. This way, it shields the data scientist from unnecessary complexities like making sure there is enough computing capacity to avoid system failures. Incidentally, if you have any questions or you run into issues getting the application working, feel free to get into contact with me directly.
Courageous commitment and resources
When drawing a personal conclusion about my experience at BMW Group IT, I could just say that I truly enjoyed working with the team and that it was a cool experience to build a fully actionable tool from scratch – especially because it is such an exciting time for machine learning in connection with software engineering. But it is actually more than that. From my point of view, the BMW Group really understands the ubiquitous need for data and how it can support the company’s value chain in order to make ultimately better decisions (and continue making cool cars and offer great services). And they act on what they say, because it takes a huge amount of upfront trust from the company-side to give a young team such levels of freedom, resources and commitment. So, thanks all! And just in case my story and experience has been catching your attention, you might be interested to know that BMW Group IT is hiring IT specialists on a constant basis.
We at BMW are developers!
If you still think we only have “fuel” in our corporate DNA, think again! As BMW Group IT, we are responsible for the entire information technology at the BMW Group – and thus a driver for the BMW Group’s digital transformation. In each of our vehicles and every service, IT plays a vital part: it is IT that provides the basis for intelligent security and connectivity. IT also enables convenient value-add services and improves sustainability. As the digital transformation is taking up speed, the role of IT is progressively growing and covers all strategic areas such as product development, technological concepts and manufacturing processes. The team at BMW Group IT takes center stage to shape the future of mobility and translate our brand’s core element „joy“ into next generation applications and innovations.
This is why BMW Group IT is excited to be part of the annual WAD congress in May. We are looking forward to connecting with people that are intrigued by the same passion for development and technologies. We are happy to share our ideas, challenges and use cases with the “WeAreDevelopers” community – whether it is in the area of Mobility Services, Q-Control, Cyber Security, IT-Services, Deep Learning or Industry 4.0. And be sure to join us for our hacking challenge and an exciting crash test driving experience!
But most importantly we are looking forward to having a good and joyful time at the congress together with you all. See you there!