At Bell, I was primarily a Python developer within the
Innovation group.
Initially, I worked with Microsoft AirSim
to create training data for object recognition machine learning.
Some of this work was featured in the Microsoft Ignite 2021 conference
during the segment on AI:
Later, I was among the first customers for Project AirSim, Microsoft’s successor
to this project. I met with Microsoft engineers weekly to discuss the product and
provide feedback. Some of this work is described in this article:
I later focused on business modeling and fleet simulation and served as the scrum
master of the team. With the help of Azure engineers, we were able to re-architect a
data analysis workflow that had been set up before I joined. Our multi-terabyte
SQL database in Azure was costly to maintain, and queries were slow.
By leveraging
Parquet files, Databricks,
and Azure Batch, we were able
to analyze data much more quickly at a fraction of the cost.
The monthly data storage cost became negligible and we were able to spin up
servers automatically on-demand to process data in parallel, and then shut them down
when not in use. This reduced our annual Azure bill by over $30,000.
I also implemented and managed nearly all of Innovation’s CI/CD with
Azure Pipelines and on-premises build agents. When I started, automated testing,
deployment, and static analysis of Python code were not in place. I established
standard development tools and settings for projects to keep code consistent and
high-quality. As demand for shared Python code grew, I published internal packages
to JFrog Artifactory, giving our team and other internal customers a controlled way
to reuse software while maintaining export compliance.
I was heavily involved in Bell’s Advanced Vertical Robotics competition. This is a
high school drone competition where teams are sent a kit of parts and follow
instructions on how to assemble, test, and fly their drone from scratch. Once they
get their drone flying, they’re expected to modify the drone to complete the challenges
given in the game.
I worked on the flight software of the drones, wrote some
documentation, and helped support students on the forums. From a software perspective,
there are a ton of moving parts. As everything is done indoors, we wrote our
own software to take data from a stereoscopic tracking camera, convert this into
fake GPS data, and then feed it directly to the flight controller. We had custom
firmware for the PX4 flight controller to facilitate this,
firmware for the onboard Arduino to manage LEDs and servos, a Python ground control
station, and all of the flight software. The flight software was made up of about 6
different software modules that all had to work together correctly. This included
code to get data from the tracking camera, code to process video data looking
for AprilTags placed on the game
court, sensor fusion code, code to interact with the flight controller, code
to interact with the Arduino, and more. We ended up with pub/sub architecture
built on MQTT where software modules were independent Docker
containers that could subscribe to, or publish, JSON data. This allowed for a very
testable design, and plug-and-play architecture that easily allowed students
to write their own autonomy code.
Finally, for seven weeks, I was a subcontractor for Raytheon, brought in to compile,
configure, and deploy their mission software to a large number of
real and virtual air-launched effects.
This was challenging, as the network connecting all of the devices was
completely isolated from the outside world. I developed a system to bundle
Raytheon’s software in a Docker container and utilized
Ansible to deploy it. This involved spending three
weeks in two separate trips to the US Army Yuma Proving Grounds for live tests.
A later test at Dugway Proving Grounds was covered in this article:
