Computer Vision Engineer

• Oct 2023 - Present

Research Assistant

• Mar 2023 - Oct 2023

• Conducted extensive literature review and analysis of hundreds of papers to understand the robustness characteristics of deep learning models and the potential defensive capabilities of ensemble models against adversarial attacks. • Developed a novel methodology for improving adversarial robustness by applying and evaluating training strategies including TRADES, SCORE, and MART, which were initially designed for individual models, on ensembles of models. • Addressed the challenge of… Show more • Conducted extensive literature review and analysis of hundreds of papers to understand the robustness characteristics of deep learning models and the potential defensive capabilities of ensemble models against adversarial attacks. • Developed a novel methodology for improving adversarial robustness by applying and evaluating training strategies including TRADES, SCORE, and MART, which were initially designed for individual models, on ensembles of models. • Addressed the challenge of lack of diversity in ensemble models by successfully incorporating ensemble diversification methods such as GAL, ADP, DVERGE, and TRS into the training process. • Implemented the complex training process in PyTorch, involving various technical optimizations including mixed precision training and parallel computation for the models, which significantly increased the computational efficiency. • Identified and rectified a key issue regarding adversarial accuracy of the ensemble models, leading to superior performance compared to single models trained with adversarial training. • Spearheaded the empirical evaluation and validation of the proposed methodology, demonstrating its efficacy and theoretical soundness. • Contributed to the broader scientific community by developing a comprehensive defense against adversarial attacks using a robust ensemble of deep models. Show less • Conducted extensive literature review and analysis of hundreds of papers to understand the robustness characteristics of deep learning models and the potential defensive capabilities of ensemble models against adversarial attacks. • Developed a novel methodology for improving adversarial robustness by applying and evaluating training strategies including TRADES, SCORE, and MART, which were initially designed for individual models, on ensembles of models. • Addressed the challenge of… Show more • Conducted extensive literature review and analysis of hundreds of papers to understand the robustness characteristics of deep learning models and the potential defensive capabilities of ensemble models against adversarial attacks. • Developed a novel methodology for improving adversarial robustness by applying and evaluating training strategies including TRADES, SCORE, and MART, which were initially designed for individual models, on ensembles of models. • Addressed the challenge of lack of diversity in ensemble models by successfully incorporating ensemble diversification methods such as GAL, ADP, DVERGE, and TRS into the training process. • Implemented the complex training process in PyTorch, involving various technical optimizations including mixed precision training and parallel computation for the models, which significantly increased the computational efficiency. • Identified and rectified a key issue regarding adversarial accuracy of the ensemble models, leading to superior performance compared to single models trained with adversarial training. • Spearheaded the empirical evaluation and validation of the proposed methodology, demonstrating its efficacy and theoretical soundness. • Contributed to the broader scientific community by developing a comprehensive defense against adversarial attacks using a robust ensemble of deep models. Show less

Development Intern

• Jul 2019 - Dec 2019

• Developed and tested a patent landscaping system that finds similar published patents using machine learning using IPC standards and by measuring contextual similarity using Bert from a database of patents with pre-generated Bert encodings. The performance of the contextual matching is improved by making a hierarchical database of patents based on IPC standard subsections in MySQL. • Designed and implemented a retail data analyzer that tracks store inventory in real-time with data gathered… Show more • Developed and tested a patent landscaping system that finds similar published patents using machine learning using IPC standards and by measuring contextual similarity using Bert from a database of patents with pre-generated Bert encodings. The performance of the contextual matching is improved by making a hierarchical database of patents based on IPC standard subsections in MySQL. • Designed and implemented a retail data analyzer that tracks store inventory in real-time with data gathered from checkout counters on a decentralized database made using Apache Cassandra and deployed on AWS. This system orders new inventory to stock the retail stores based on the current and expected stock of items. • Developed a parking allotment system that scans the number plates of the entering vehicle by detecting OCR. Then it allots a parking spot based on the current occupancy is noted using image processing on an eagle-eye view of the parking lot. this solution has been tested and validated to work at multiple locations. • Developed and implemented machine learning models for predictive analytics, customer segmentation, and anomaly detection. • Processed and analyzed large datasets using Python, SQL, and Spark. • Collaborated with stakeholders to identify business problems and translate them into data-driven solutions. • Conducted experiments to evaluate the performance of machine learning models. • Communicated findings and recommendations to non-technical stakeholders through data visualizations and presentations. Show less • Developed and tested a patent landscaping system that finds similar published patents using machine learning using IPC standards and by measuring contextual similarity using Bert from a database of patents with pre-generated Bert encodings. The performance of the contextual matching is improved by making a hierarchical database of patents based on IPC standard subsections in MySQL. • Designed and implemented a retail data analyzer that tracks store inventory in real-time with data gathered… Show more • Developed and tested a patent landscaping system that finds similar published patents using machine learning using IPC standards and by measuring contextual similarity using Bert from a database of patents with pre-generated Bert encodings. The performance of the contextual matching is improved by making a hierarchical database of patents based on IPC standard subsections in MySQL. • Designed and implemented a retail data analyzer that tracks store inventory in real-time with data gathered from checkout counters on a decentralized database made using Apache Cassandra and deployed on AWS. This system orders new inventory to stock the retail stores based on the current and expected stock of items. • Developed a parking allotment system that scans the number plates of the entering vehicle by detecting OCR. Then it allots a parking spot based on the current occupancy is noted using image processing on an eagle-eye view of the parking lot. this solution has been tested and validated to work at multiple locations. • Developed and implemented machine learning models for predictive analytics, customer segmentation, and anomaly detection. • Processed and analyzed large datasets using Python, SQL, and Spark. • Collaborated with stakeholders to identify business problems and translate them into data-driven solutions. • Conducted experiments to evaluate the performance of machine learning models. • Communicated findings and recommendations to non-technical stakeholders through data visualizations and presentations. Show less

Machine Learning Intern

• May 2018 - Jul 2018

• Developed and tested a Product classification system that classifies products using their name and description into their respective UNSPSC codes using a hierarchical classification model that classifies into the lower subsection on each level of classification. The name and descriptions are vectorized using Word2Vec. • Developed and maintained ML Infrastructure and ML pipelines, which resulted in increased efficiency of model training and deployment. • Extended the existing ML Platform… Show more • Developed and tested a Product classification system that classifies products using their name and description into their respective UNSPSC codes using a hierarchical classification model that classifies into the lower subsection on each level of classification. The name and descriptions are vectorized using Word2Vec. • Developed and maintained ML Infrastructure and ML pipelines, which resulted in increased efficiency of model training and deployment. • Extended the existing ML Platform and frameworks for scaling model training and deployment. • Partnered closely with various business and engineering teams to drive the adoption, integration of model outputs. • Implemented MLOps using Databricks in the Cloud, leading to improved performance, reduced downtime, and faster deployment times. • Developed and maintained Python scripts for data analysis, model building, and deployment. • Used Spark and SQL to process large datasets for model training. • Contributed to the company's ML ecosystem by sharing best practices, training team members, and mentoring junior developers. Show less • Developed and tested a Product classification system that classifies products using their name and description into their respective UNSPSC codes using a hierarchical classification model that classifies into the lower subsection on each level of classification. The name and descriptions are vectorized using Word2Vec. • Developed and maintained ML Infrastructure and ML pipelines, which resulted in increased efficiency of model training and deployment. • Extended the existing ML Platform… Show more • Developed and tested a Product classification system that classifies products using their name and description into their respective UNSPSC codes using a hierarchical classification model that classifies into the lower subsection on each level of classification. The name and descriptions are vectorized using Word2Vec. • Developed and maintained ML Infrastructure and ML pipelines, which resulted in increased efficiency of model training and deployment. • Extended the existing ML Platform and frameworks for scaling model training and deployment. • Partnered closely with various business and engineering teams to drive the adoption, integration of model outputs. • Implemented MLOps using Databricks in the Cloud, leading to improved performance, reduced downtime, and faster deployment times. • Developed and maintained Python scripts for data analysis, model building, and deployment. • Used Spark and SQL to process large datasets for model training. • Contributed to the company's ML ecosystem by sharing best practices, training team members, and mentoring junior developers. Show less