Beyond the Model: Building Real-World Machine Learning

In the latest Developer Impact Series, Dave Neary of Ampere® Computing talks with Dr. R.J. Nowling from the Milwaukee School of Engineering to discuss how the school is bridging the gap between theoretical machine learning (ML) and real-world production systems and about how students learn to build “real” ML systems that work in everyday software — not just clever math models.

Dr. Nowling explains that lots of schools teach students how to design smart computer models. MSOE tries to go one step further. The key idea is: a model isn’t helpful if it only works in a lab. In the real world, a working system also needs other parts — like getting fresh data, connecting to databases, turning raw information into the kind of numbers the model can use, and keeping track of whether the model is still doing well over time.

He compares it to his past industry work in online advertising. There, the model itself was only a small part of the whole process. The system had to constantly bring in new data, update information regularly, and check the “most recent” information because that’s often the best for making quick decisions. Also, models can “drift” and become less accurate as the world changes, so the system must retrain and update the model periodically. That larger set of tasks is what makes machine learning production work difficult — and important.

In his course, students practice building the full system from beginning to end. They set up databases (like PostgreSQL), create data pipelines, train models offline, and then launch them as services so that other software can use them. They also learn to monitor the model, notice when performance changes, and use automated tools so the model gets retrained and redeployed when needed. A big teaching goal is to let students see real operational problems, such as running out of disk space or having container problems, so they learn how to fix issues as engineers do in the real world.

The video also explains why Ampere computers are part of the training. Dr. Nowling says they run many student lab environments on a single Ampere server, using CPU power rather than relying on GPUs. Students typically share the machine through virtual computers, so they can each run their own setup — databases, pipelines, services, and automation jobs — at the same time. This is important because many people assume “AI always needs a GPU.”

Dr. Nowling says that’s not always true. For many practical tasks — like spotting fraud quickly or deciding whether to show an ad — speed matters and CPU-based approaches can work well. The school wants students to understand those real-life constraints, not just the most famous AI setups. He also talks about how the software works smoothly on this system. They use a common Linux operating system (Debian) and tools to set everything up, so students don’t get stuck fighting software issues.

Finally, the program is also expanding into generative AI (like tools that can produce text). Students learn that real-world AI systems often combine different “smarter but smaller” models instead of using one huge model for everything, which can reduce cost and improve reliability.

Watch the full video here:

To learn more about Ampere’s developer efforts and find best practices, visit Ampere’s Developer Center and join the conversation in the Ampere Developer Community.

Check out the full Ampere article collection here.