I worked at Amazon as an SDE I in the Forecasting Deep Learning team.

My team owned quite a few forecasting models which ingest large amounts of raw data to output forecast bundles for our internal customers to consume. We worked with multiple pipelines for various models, which kicked off multiple workflows every day that generated these forecasts. Our customers varied from the data science side of our team who analyzed the forecasts to make sure nothing is amiss, to more downstream teams that took in our forecasts and transform and compile them into more consumable formats for different stores to utilize. Our team worked on a sprint system, and I generally switched between different projects as sprints start and end. I did a variety of different tasks depending on what was needed, but I’ll list some of the more notable/frequent ones.

My most recent project involved designing and implementing a tech stack consolidation between two of our more prominent models. The previous year, our team had taken ownership of quite a few core forecasting models that other teams had owned during a re-org, and one of the larger models we inherited had a fairly large and complex code base that was difficult to deep dive or upkeep. The goal of my project was to both re-organize how the model workflow worked to become more efficient, as well as reformat it in a way that it mimicked the structure of our main pre-existing model before the re-org. The design revolved around deciding how we wanted to allocate the different processes in the model, and figuring out where we wanted to utilize the computing resources of EMR clusters and how we wanted to implemeent external APIs to simplify the process.

Some of my work consisted of modifying or creating new datasets depending on what features our customers wanted. Our workflows essentially flowed the input data through these various datasets to filter and calculate information needed for our models to run properly. These datasets need to be able to process large amounts of data quickly and efficiently in order for our workflows to generate forecasts in a timely manner. Modifying the datasets is usually made for the purpose of either removing unnecessary steps in the workflow or changing the way its run so that the workflow runs more efficiently.

Another main task involved thorough maintenance and updates to our pipelines and workflows. As other teams update certain inputs or schemas for their own improvements, we need to backtest and forward test many components of our own workflows in order to keep up to date and make sure the forecast accuracy doesn’t go down in the process. These workflows utilize a variety of different amazon services including (but not limited to) AWS, EMR, S3, Sagemaker, and Athena. I’ve familiarized myself with quite a few amazon services over the course of my time here.

Leica Geosystems focuses its efforts on producing products and systems for geographical measurement and surveying.

My main goal as a software development intern was to create a C# program that essentially demo-ed most of the features that Leica’s software do. It took in Cyclone, Jetstream, and LGS files (all 3D point cloud manipulating software) and allowed for manipulation of the point cloud. I was to base all of the C# features on a similar C++ program that was already implemented. The goal was for my program to be used by other C# developers as a basis to see how Leica’s products worked. Throughout working on this project, I had spent ample time learning debugging techniques for both C# and C++ in Visual Studios, as well as learned how to share information to and from C++ native code to the managed C# code. I worked with callback functions and many kinds of typecasting, and learned how to work with much larger projects than what I was used to (over tens of thousands of lines of code, and many different classes and files I had to understand in order to implement the program). The general outline of the program was there, such as the base classes of 3D lines, quaternions, cylinders, etc. but most of the tools to utilize them had to be implemented by me. I also gained experience in documentation as well, documenting the api of the C# program for future developers to use.