Beyond full-stack: The rise of the ML product engineer

At Samsara, we’re looking for a new kind of full-stack engineer: the ML Product Engineer. These are engineers with product understanding, infra knowledge, and ML skills to build exactly what our customers need to run their organisation better. We have been building intelligence all through our systems, in big and small ways, and now we are giving power to the engineers who know the most about the product, so they can directly use ML to improve it. 

Our custom hardware is installed on millions of vehicles, we process over 25 trillion data points annually, and we are aggressively using AI to improve driving safety, improve fuel efficiency, and solve fleet-wide routing and navigation. Sometimes we are doing this in the cloud, and often we are moving the intelligence to our hardware on the edge: sensors, cameras and other devices that integrate together to give a full picture of a fleet’s connected equipment. 

We are building these muscles internally, and also looking to hire top-notch makers and builders who know how to use ML and world-class infra to solve end-to-end problems.

Managing anomaly detection

Whether you’re protecting drivers from security threats or preventing fuel theft, having a clear view of your operations is essential. We listened to organisations in Mexico about their most pressing safety concerns, from carjacking to fuel and asset theft. By listening to our customers around the globe, we’ve found that many need to know the moment a vehicle makes an unscheduled stop, when fuel levels don’t match real-world driving, or if equipment has gone completely off route and possibly stolen. 

So we launched multiple initiatives to identify these problems, working with our internal teams of ML experts. With their expertise being in high demand across the organisation, we recognised that scaling these innovations required a new approach. After embedding their ML experts in the team, we found that our engineers were excited to learn how to do this themselves and get more hands-on with cutting-edge techniques. We want to add AI and ML everywhere in our product, and we realized that the fastest way to scale this up is to give that power directly to the engineers.

Our core product leverages the Go language heavily, both on the firmware and the infra side, even in our core data engineering pipelines, but most ML development is done in Python. We found that there was a big disconnect between getting the core Python ML algorithms working and then moving feature generation into Go so we could deploy them, and this added a lot of delay. To speed this whole process up, we worked with the ML Infra team to create a self-service pipeline where both ML and product engineers could prototype in Python notebooks and then use Spark Structured Streaming to directly integrate with our Go infra, our Kinesis-based streaming system, and our Databricks-based data lake.

The Power of ML Product Engineers 

We have found that product engineers can create classifiers and prediction algorithms to move quickly and add amazing new features to our products, leveraging their core understanding of our customers’ needs. We are leveling up our internal talent, which has years and sometimes decades of experience in traditional engineering, and teaching them how to curate large datasets, define metrics and loss functions, and experiment with the wide variety of machine learning algorithms.

Of course, for the really deep, cutting-edge work, we still use dedicated ML experts who are brilliant at creating, tuning and deploying models, which the product engineers then connect with core product functionality. But there are countless small (and medium-sized) ways we can improve the customer experience, where our product experts can now move to immediately improve the product. And if the product engineers find that it’s a tougher problem than expected, they can now work side-by-side with the ML experts to solve it, faster than ever, because they are learning to speak the same language. 

We have multiple initiatives shipping in the next few months using this new, collaborative model to ship products for fuel theft, fuel card fraud detection, and unexpected time or location of equipment. Please reach out to us if you’d like to join us on this journey.