Vivekjyoti Banerjee

B.S. in Computer Science (Machine Learning Specialization)

B.S. in Mathematics (Statistics Specialization)

Aug 2018 - Dec 2021

University of Maryland, College Park, MD

• GPA: 3.723 / 4.0
• Quality Enhacement Systems and Teams (QUEST) Honors program student
• Outstanding Service Award, QUEST Honors Program
• 5x University of Maryland Dean's List
• Maryland Delegate Sydnor Merit-based Scholarship
• Selected Student Speaker at the Annual QUEST Conference for Cohort 34

Machine Learning Engineer

Mar 2022 - Present

Carnegie Robotics LLC, Pittsburgh, PA

Minefield Hazard Detection

• Improve neural networks for military applications to recognize a variety of buried hazards using Ground-Penetrating Radar (GPR) signals and traditional signal processing techniques. This improves the performance of state-of-the-art detection models by over 10%.
• Rewrite ResNet backbone to create novel fusion block, combining 2d & 3d convolutions on volumetric data.
• Develop a 7-stage automated, parallelized, and reproducible pipeline to filter GBs of sensor data and correct inconsistencies, tripling dataset size while reducing data generation time from 30+ hours to 2 hours.
• Reconstruct receiver end-effector positions for a 5-degree of freedom arm to improve and create dense radar data representation for machine learning.
• Run ablation studies with multiple backbones, fusion blocks, and data construction methods to uncover the best algorithm for single and multi-class hazard detection
• Write custom network analysis tools to explain inference results

Manifold Visualizer

• Compare self-supervised contrastive learning vs. momentum encoding vs. self-distillation with no labels frameworks for image retrieval tasks on proprietary data.
• Redesign and improve ResNet and vision transformer-based networks to generate embedding spaces.
• Design and iterate on a 4-stage parallelized multi-model inference pipeline to generate embeddings for 900+logs on EC2 using LogQS (cloud storage system for robotic logs) with 50x speedup.
• Implement an interactive tool to visualize and explore 3D embeddings spaces of various datasets with Dash.
• Generalize visualization tools to be integrated and used for a variety of detection and segmentation tasks.
• Manage an intern, create, and delegate tasks, and review intern git pull requests.

Software Engineer - Computer Vision

Aug 2021 - Dec 2021

University of Maryland, College Park, MD

• Applied Computer Vision techniques to evaluate sports players' performance and improve technical skills
• Redesigned the existing neural network pipeline to compute velocities of moving objects resulting improved accuracy from 50% to 92%
• Designed backend infrastructure to measure players' technique, power, and accuracy using methods from published research papers
• Compiled Python scripts to streamline data generation by labeling keypoints and bounding boxes from video footage to work across iOS and Android devices

Machine Learning Intern

Jun 2021 - Jul 2021

Bechtel, Reston, VA

Big Data Analytics Center

• Research and apply SotA document segmentation networks for engineering blueprints and table detection to digitize proprietary data.
• Experiment with transfer learning with a Res-Net model fine-tuned on company documents to achieve 98% detection accuracy.
• Develop synthetic data to leverage the benefits of an annotated dataset to improve accuracy by 10% on scanned data.

Information Systems and Technology

• Implemented a rule-based system using entity resolution to insure complete and consistent data handover to clients
• Reduced data validation time from 4+ hours to a few seconds using Python scripts to verify CFIHOS standards
• Conducted research on drone imagery and photogrammetry software use cases to develop digital twins for improved project planning and maintenance

Machine Learning Researcher

Dec 2020 - Dec 2021

University of Maryland, Center for Automation Research

• Design controlled experiments to differentiate twin, fraternal, and sibling faces and measure human accuracy against a neural network developed by the lab during the IARPA Janus program.
• Evaluate networks and humans’ performance on a variety of viewpoints to conclude that the network performs at the level of the best humans in the study or outperforms them with statistical significance.
• Reduce 512-dimensional face feature vectors using t-SNE and principal component analysis to identify similar face clusters used in experiment data.

Quantitative Analyst

Sep 2020 - Dec 2020

Bechtel and QUEST, College Park, MD

• Evaluated time series data from 1992 to 2019 on bridges across five U.S. regions
• Processed over 6 million bridge entities per year to develop a training data set
• Applied dimensionality reduction based on empirical approaches to build a robust model with 94% accuracy
• Built regression model to forecast emerging hot-spots for bridge reconstruction in the next five years

Enrichment Program Teacher

Oct 2018- Sep 2020

Coder Kids Inc., McLean, VA

• Showed the basics of 3D printing to students (6-8)
• Taught fundamentals of Scratch and Python principles of programming to elementary and middle school students
• Helped students create mesh models to 3D print for design competitions.