Experience

Casetext

• United States
• Technology, Information and Internet
• 1 - 100 Employee

• Senior Machine Learning Researcher

  • Jan 2022 - Present

  Design, develop, and deploy scalable solutions for legal and transactional natural language understanding, question answering, and retrieval Designed, developed, and manage large-scale test suite for evaluating model and prompt changes to CoCounsel, Casetext's AI legal assistant based on OpenAI's GPT-4 Pretrain and fine-tune Transformer-based vector retrieval and cross-encoder models and integrate them with generative AI models in order to improve overall system speed, answer… Show more Design, develop, and deploy scalable solutions for legal and transactional natural language understanding, question answering, and retrieval Designed, developed, and manage large-scale test suite for evaluating model and prompt changes to CoCounsel, Casetext's AI legal assistant based on OpenAI's GPT-4 Pretrain and fine-tune Transformer-based vector retrieval and cross-encoder models and integrate them with generative AI models in order to improve overall system speed, answer quality, and hallucination rate Design and oversee data annotation process for legal training data Show less



Oak Ridge National Laboratory

• United States
• Research Services
• 700 & Above Employee

• Research Scientist

  • May 2021 - Jan 2022

  Primary technical lead of 10+ member research team working on multi-year collaboration with National Cancer Institute to automate cancer surveillance using deep learning methods Delivering production-ready automated pathology report coding API into NCI SEER registries, directly resulting in 15%-25% workload reduction for human registrars Developing novel algorithms for knowledge-oriented pretraining of Transformer language models on long clinical and biomedical… Show more Primary technical lead of 10+ member research team working on multi-year collaboration with National Cancer Institute to automate cancer surveillance using deep learning methods Delivering production-ready automated pathology report coding API into NCI SEER registries, directly resulting in 15%-25% workload reduction for human registrars Developing novel algorithms for knowledge-oriented pretraining of Transformer language models on long clinical and biomedical text Developing capabilities for massively distributed pretraining of custom Transformer architectures on Summit and Frontier leadership-class supercomputers

• Deep Learning Postdoctoral Researcher

  • Jan 2020 - Apr 2021

  Knoxville, Tennessee Area Develop and deploy deep learning architectures for automated information extraction from cancer pathology reports; techniques include CNN, RNN, and Transformer-based approaches Develop autonomous driving agents in CARLA simulation environment using end-to-end deep imitation learning and deep reinforcement learning methods Develop and apply multi-task, transfer learning, semi-supervised, visualization/interpretability, and uncertainty quantification methods for deep learning… Show more Develop and deploy deep learning architectures for automated information extraction from cancer pathology reports; techniques include CNN, RNN, and Transformer-based approaches Develop autonomous driving agents in CARLA simulation environment using end-to-end deep imitation learning and deep reinforcement learning methods Develop and apply multi-task, transfer learning, semi-supervised, visualization/interpretability, and uncertainty quantification methods for deep learning models Scale deep learning algorithms across multiple GPUs and nodes on Oak Ridge supercomputer clusters

• Deep Learning PhD Student

  • Jan 2017 - Dec 2019

  Developed new state-of-the-art text classification model for cancer pathology reports based on neural self-attention; the approach achieves better accuracy and trains over 10x faster than the previous state-of-the-art method Developed visualization tool for interpretable deep learning for clinical text classification using neural attention weights Developed novel methodology to identify and correct for mismatches between human expert annotations and the content reported within… Show more Developed new state-of-the-art text classification model for cancer pathology reports based on neural self-attention; the approach achieves better accuracy and trains over 10x faster than the previous state-of-the-art method Developed visualization tool for interpretable deep learning for clinical text classification using neural attention weights Developed novel methodology to identify and correct for mismatches between human expert annotations and the content reported within individual cancer pathology reports; this method improves classification accuracy on cancer pathology reports by up to 10%



The University of Georgia

• Higher Education
• 700 & Above Employee

• Graduate Research Assistant

  • Aug 2016 - Dec 2016

  Work with interdisciplinary team on human activity recognition project that attempts to classify activity type based on hip-worn accelerometer device Developed convolutional-LSTM model that achieves competitive performance on human activity recognition tasks without requiring manual engineering of features



Noble Systems

• United States
• Telecommunications
• 100 - 200 Employee

• Technical Writer

  • Mar 2012 - Jul 2016

  Greater Atlanta Area Produce customer-facing online training for a wide range of contact center products, including campaign management software, IVR scripting interfaces, and more Maintain and develop structure, templates, procedures, and single-sourcing guidelines for internal, VAR, and customer knowledge bases—content includes product technical specifications, client connectivity information, troubleshooting and configuration guides, database reference tables, and best practices Troubleshoot all… Show more Produce customer-facing online training for a wide range of contact center products, including campaign management software, IVR scripting interfaces, and more Maintain and develop structure, templates, procedures, and single-sourcing guidelines for internal, VAR, and customer knowledge bases—content includes product technical specifications, client connectivity information, troubleshooting and configuration guides, database reference tables, and best practices Troubleshoot all technical problems related to internal and customer knowledge bases, including issues with HTML/CSS formatting, Team Foundation Server version control, and nightly auto-build and publishing process Show less