As a software engineer at Verdant Robotics, I have had the opportunity to work on a variety of exciting and challenging projects. One of the main accomplishments was the transition from a propriety NVIDIA Tao toolkit to a Pytorch-based model pipeline, allowing for faster adaptation of new model architectures. I also introduced new model architectures and techniques for focusing training efforts on high-impact labels, reducing training time and costs for new crops.

In addition, I developed an asynchronous data extraction platform for automated image extraction from production data, reducing the time of the ingestion process. I introduced dataset visualization processes using FiftyOne and created visualizations and statistics for better insight into the distribution across multiple criteria using Pandas, Plotly, and sklearn.

I maintained the central evaluation backend used for reporting model performance across the organization and introduced Python tools for best practices. I improved the model deployment process, including performance benchmarks for new model architectures, and introduced semi-supervised training approaches based on Vision Transformer for a classification project I executed. I also introduced model calibration techniques for the classification model and was in charge of the project to create superresolution images using a generative adversarial network.

As a Computer Vision Engineer at Dexterity, Inc., I contributed to the development of a robot singulation station that was deployed on multiple sites at one of the top 3 logistics companies, leading to multiple patents in the field. In this role, I was responsible for various aspects of the computer vision system, including the adoption of a new camera model across different product lines, and the development of a computational node to merge outputs of multiple cameras together based on an external camera calibration process with the help of Charuco markers.

I also conceptualized and implemented a Data Collection System for pre-deployment on customer facilities to bootstrap data collection for faster deployment. Additionally, I worked on a system based on RGBD cameras to flag misplacement of parcels and avoid robot collision on a high throughput conveyor belt. This included the development of a classification model to flag wrong robot actions. I was also in charge of the second generation of the Computer Vision pipeline based on object detection which resulted in a higher accuracy, a reduction of the system complexity, and an error reduction.

In addition to these responsibilities, I introduced MLFlow and WandB for model management and enhanced reporting for the broader organization. Finally, I built a label management toolchain for fast in-house labeling to speed up research on new methodologies.

Most notably, I developed a flow detection system that dramatically reduced the number of miss picks of the system and led to higher throughput. My contributions helped to increase the accuracy, efficiency, and overall performance of the system, resulting in successful deployments at multiple customer sites.

At Dexterity, Inc., I worked on several projects related to robotics and computer vision, including developing the Python client for the visualization stack, detecting car door handle orientation using classical computer vision algorithms, and conducting a camera comparison study between different models.

My internship allowed me to apply my knowledge in a real-world setting and further develop my skills in robotics, computer vision, and Python programming.

As a student assistant at the Karlsruhe Institute of Technology, I worked on the backend of a system that allowed students to submit their projects online for execution in the KIT robot lab. This project was done in collaboration with Udacity. Through this role, I gained valuable experience in software development, project management, and teamwork.