This is the first part of a two part series on my first co-op. This part is about the work: what I built and what I'm doing next. The second part is about living in and travelling around Europe.

For my first co-op I worked in Stuttgart, Germany, at the German Aerospace Center (DLR), Institute of Vehicle Concepts. I earned an outstanding rating. Here's what the work was and how I did it.

Euro 7 is the seventh set of European regulations on vehicle emissions. It's the first to set a hard limit on non-exhaust emissions, meaning the particles that come off tires and brakes rather than out of the tailpipe. Measuring those in a standardized way is still an open problem. The regulations take effect in November 2026, and my institute was working on them ahead of that deadline.

The question we were trying to answer: how do you measure particle emissions from both ambient and mobile data sources? The main intern project was a website visualizing particle emissions from a moving vehicle and a fixed sensor location, built to demonstrate our approach internally and to project partners.

My two supervisors, both master's/PhD researchers, were busy with other projects, so I had a lot of freedom in how to solve it. I worked alongside another intern, so most of this was a team effort.

Our first implementation used deck.gl, 3D animations, and CSV/GLB files, built around smooth scrolling and heavy visuals. It demonstrated the solution but never made the use cases clear, so we went back to the drawing board and presented the same visualizations more simply.

Around this time my supervisors started trusting me with other work. I got most interested in the data pipeline from car to cloud.

The first side project was a Raspberry Pi 4 human machine interface (HMI): a dashboard mounted in the research vehicle's cup holder. It published and subscribed to two MQTT topics, one for telemetry and one for commands, and let the test driver start and stop collection from the sensor node.

This was a necessity. The sampling hardware is sensitive to weather, and the driver needed a way to protect it mid-drive without ending the run.

The HMI also showed live sensor readings from the node during the drive. I tested this myself on the autobahn in Stuttgart. Accelerating hard to 150km/h or braking at a stop sign moved the readings immediately.

Field hardware has to survive weather and patchy coverage, so I wrote a reconnection watchdog that handled dropouts and unclean shutdowns. Running an ethernet cable wasn't possible without heavily modifying the setup, so everything had to work over cellular.

I'm proud of this one. By the end of my internship the HMI had been field tested over thousands of kilometres and hundreds of driving hours. It let the researchers collect far more observations, and it kept the hardware out of trouble more than once.

The second side project was hardware. My supervisor wanted the existing Arduino/PlatformIO code ported to Zephyr RTOS. It took much longer than expected. I finished working code, but between hardware issues and limited time on the bench it was only half done: it worked in software, but there wasn't time to validate it on hardware before I left.

I learned a lot of Zephyr along the way, including the device tree, config, and the T2 workspace layout. I wouldn't say I have it down. I just learned it, and it was a great experience.

The third project was a set of Grafana plugins for the researchers: a 3D hexagonal drive heatmap, a 3D ambient emissions heatmap, a particle size distribution graph, a 3D version of that graph, and a drive selector.

The hard part was supporting two InfluxDB versions at once. Every plugin had to be modular: fields and buckets differ between 1.x and 2.x, and 1.x uses InfluxQL while 2.x uses Flux, so the same data arrives at the plugin in different shapes. On top of that, the plugins had to read time series data and stay flexible about which fields they took in.

The 3D hexagonal heatmap showed this best. Mostly the researchers wanted readings from the vehicle plotted over position and time. But they also wanted the same drive against vehicle speed, battery state of charge, and my favorite, cellular signal across the route.

The position data was precise enough to follow the exact path of a drive. You could also watch the network drop out, with clear coverage gaps along the route. This heatmap became the lead visual for the project.

It was useful to the researchers too. Without the size distribution graph they had to parse values by hand and write a MATLAB script for a basic plot. With it they could read results live during testing and reach conclusions faster.

The fourth and last side project was a CAN bus signal parser handling 11-bit and 29-bit IDs, extending an existing parser. I learned to read standard OBD-II messages on one vehicle, then applied the same approach to another.

This was my favorite part of the internship. It felt like everything else had led up to it, because the signals fed straight into the pipeline I'd spent months improving. Adding vehicle speed to the parseable signals meant it moved through the I2C Zephyr sensor bus I built, appeared on the Pi HMI, and showed up as a new option in every Grafana plugin I'd written.

I also got to field test a second vehicle, which was a nice change of pace.

By August, with the side projects done, the main intern project came together. We'd been through several iterations of the landing page and the overall design, and settled on something minimal that explained the solution, the story behind it, and what we offered, without burying the visualizations under Grafana dashboards and technical detail.

Two moments stand out. The first was shooting b-roll of the vehicle. The other intern has a gimbal and a good camera, and we'd been joking at work about how nice it would be to have real footage of the car for the site. We asked our supervisors and they were immediately on board. We planned the scenes ahead of time: close-ups, drivebys, and shots of someone using the HMI.

The second was a script I wrote to cut page latency. The site had heavy CSV and GLB animations, so I stepped through the deck.gl animations, screenshotted each frame, and stitched the PNGs into a seamless video with ffmpeg. It cut load time and page size, and it meant the site didn't have to serve the large 3D asset or our raw CSV data. I later reused the technique to speed up the UWHPC loading animation.

All in all it was a fulfilling internship. All of my co-workers spoke fluent English, and I'll miss the outings to Pforzheim and the beach volleyball.

My supervisor offered me a return and said I was welcome to come back. I turned it down for two reasons. The first is practical: I can't get a second Youth Mobility Visa for Germany. The second is money. The compensation covered living costs in Stuttgart, but there wasn't enough left over to save for next term or put toward tuition. So I'm looking in Canada for my next co-op.

The work itself gave me one of the best experiences I could have asked for: breadth and depth across hardware, firmware, fullstack, infra, a bit of ML, and a lot of hands-on testing, deployment, and system design.