Aug 2024 – Mar 2025 · 7 mos · New York City Metropolitan Area · Remote

• Spearheaded the development of a machine learning, data pre-processing & analysis pipeline using AWS, employing Agile methodologies for efficient and iterative progress.
• Incorporated ONNX for efficient model deployment, achieving 96.9% accuracy with a classification model developed on balanced skewed data, significantly reducing manual labelling costs.
• Attained a 0.47 mAP with YOLOv8 object detection, expanding project scope to identify and categorize diverse architectural elements.

Jan 2023 – Dec 2023 · 11 mos · Boston, Massachusetts, United States · On-site

• Worked under supervision of Prof. Eshed Ohn-Bar improving deep learning models.

Jan 2023 – May 2023 · 4 mos · Boston, Massachusetts, United States · On-site

• Teaching Assistant for Data Science Tools and Application course (CAS CS 506) at Boston University.
• Assist in designing and delivering lectures and tutorials in Data Science and Machine Learning.
• Provide hands-on support to students during lab sessions and practical exercises.
• Evaluate student assignments and provide constructive feedback to help them improve.
• Mentor and support students in their projects and research work.

Aug 2021 – Jul 2022 · 11 mos · Banglore · Hybrid

• Improving Domain Adaptation through Class Aware Frequency Transformation (IJCV '23)
• Utilized frequency transformation methods to reduce domain shift between source and target domains, using PyTorch.
• Designed an unsupervised learning approach that employed Gaussian Mixture Model to filter out clean and noisy labels.
• Achieved a 6.5% improvement in accuracy on domain adaptation datasets (Office-Home, VisDA).
• CoNMix for Source-free Single and Multi-target Domain Adaptation (WACV '23)
• Proficient in source-free domain adaptation, adapting machine learning models to new tasks without labeled data by identifying domain similarities and differences, enabling effective and efficient knowledge transfer.
• Designed a novel framework to reduce domain shift where source data is unavailable due to privacy restrictions.
• Nuclear-norm maximization technique to improve class-discriminability and prediction diversity, in conjunction with consistency between label-preserving augmentations and pseudo-label refinement methods for reducing noisy pseudo labels, resulting in significant improvements.
• Empirically showed that Vision Transformers are more suitable for domain adaptation tasks than Convolutional Neural Networks, & demonstrated that our suggested loss functions perform better on Vision Transformers.
• Holistic Approach to Measure Sample-level Adversarial Vulnerability (CVPR-W '22)
• Quantified adversarial vulnerability of a sample by combining the degree of model’s reliance on high-frequency features and the sample-distance to the decision boundary.
• Demonstrated the limitations of individual measures in characterizing inherent bias associated with adversarial attacks, and proposed a holistic metric for a more accurate assessment of model vulnerability across different subgroups.

Apr 2020 – May 2021 · 1 yr 1 mo · Remote

• Designed and developed an efficient multi-agent reinforcement learning system using Voronoi diagrams to optimize workload distribution and facilitate computation and representation of the search space.
• Implemented the Asynchronous Advantage Actor-Critic (A3C) algorithm in Pytorch to optimize the learning of agents and improve their performance in the task.
• Conducted extensive experimentation and analysis to validate the effectiveness of the proposed approach, demonstrating strong skills in critical thinking and problem-solving.

Jul 2019 – Sep 2020 · 1 yr 2 mos · Nagpur, Maharashtra, India

Apr 2019 – Jul 2020 · 1 yr 3 mos · Nagpur, Maharashtra, India

Jun 2018 – May 2021 · 2 yrs 11 mos · Nagpur, Maharashtra, India · On-site

Feb 2018 – Mar 2018 · 1 mo · Nagpur, Maharashtra, India