Jacob Krucinski

I am currently a second-year Master's student at Northeastern University studying Artificial Intelligence. I graduated from the University of Connecticut in May 2024. I am passionate about machine learning and computer vision. Through my internships, coursework, and personal projects, I have developed cross-disciplinary skills in computer science, electrical engineering, and mechanical engineering. My goal is to pursue a career as a machine learning engineer for physical systems and solvng complex problems such as autonomous driving and surgical robotics.

Email  /  Resume  /  Latest CV  /  LinkedIn  /  Google Scholar  /  GitHub  /  StackOverflow  /  Photography

My current research interests lie in the intersection of machine learning and computer vision in physical systems, along with natural language processing.

As a member of the Silicon Synapse Lab, I will be investigating various real-time Electronic Image Stabilization (EIS) techniques for a flapping-wing robot, as this lightweight platform isn't capable of carrying a gimbal or other stabilization hardware.

As a member of the Aits Lab, I explored key models such as BioGPT and SciBERT for relation extraction in biomedical literature using annotated corpora including BC5CDR, ChemProt, and DrugProt. Specifically, I used a combined ChemProt and DrugProt corpus for SciBERT fine-tuning and improved precision, recall, and F1 score by 5%.

This paper presents computationally efficient CNN models for detecting and localizing streaking targets from an optical sensor's Focal Plane Array. When compared to state-of-the-art (SOTA) probabilistic techniques, the machine learning models were 340 times faster for detection and approximately 360 times faster for localization on full size 256x256 images. For detection, the ROC AUC was 0.94 for the ML model and 0.85 for the SOTA model. For localization, the MSE was 0.0351 for the ML model and 0.243 for the SOTA model.

This project entailed developing a dynamic hand gesture recognition system from a webcam feed for computer control. The goal was to enable intuitive interactions without the use of keyboard and mouse that mimic common touchscreen/tablet gestures. Our system recognizes scrolling up/down, zooming in/out, and switching applications left/right. For this project, we developed all stages of the ML workflow, including data collection, model training/validation, and deployment interface. It can be reused for support of new gestures and computer controls.

The goal of the project is to provide beginner bicyclists with an easy learning experience when first using a CVT equipped bike, and to ensure a more efficient ride for advanced cyclists. In collaboration with Transcend Bicycle, UConn CSE Team 11, and ME Team 50, I developed the hardware and software infrastructure for the CVT controller that would perform gear shifts automatically. This involved the use of Simulink to model the Human-Bike-CVT system and test the controller in various real-world scenarios, as well as the Embedded Coder Toolbox to automatically generate C code that was deployed on our microcontroller.

The goal of this project was to build a fixed-base two DOF articulated manipulator that would play Cup Pong. The manipulator was built with 2x Dynamixel MX64-AR motors and supporting mounting hardware, as well as a Lego NXT motor for turret actuation. The microcontroller used was a Raspberry Pi 4. The camera used was a Raspberry Pi Camera Module v2.1. For cup detection, 100mm 6x6 ArUco tags were used and distance was calculated via the calibrated camera using OpenCV.