• Developed AI system to detect mental health and cancer symptoms in clinical review charts, identifying veterans needing care. • Led ORNL-VA collaboration, processing 854,543,844 rows in 136 hours (1,748 rows/sec) using SQL server and HPC cluster in a secure offline environment with a combination of PySpark and multi-threaded SQL. • Developed "prompt-ception" technique, using AI to create prompts for other AI models, improving overall accuracy, combining Claude 3.5 and GPT-4o outputs for optimal accuracy. • Implemented open-source LLaMA 3.1 70B model, achieving 94% accuracy (within 1% of GPT-4o) on 13,628 rows in 19 hours with an custom PyTorch NVIDIA CUDA environment. • Executed dictionary lookup at 33.5M words/second, demonstrating exceptional optimization skills. • Managed secure data transfer and processing in high-security offline environment. • Delivered technical presentations to 216+ stakeholders, effectively communicating complex AI concepts and latest prompting techniques for medical note classification. • Overcame Azure GPU failures, developing on-premises solutions to meet project deadlines by leveraging on-premises GPUs for high-performance AI computations, achieving significant cost savings over cloud solution.

- Established standard operating procedures as a Data Scientist and Scrum Master to guide effective data science collaborations, ensuring consistency and efficiency in project delivery.- Conducted exploratory data analysis on Long COVID data, uncovering valuable insights into incidence and prevalence rates across veteran populations to inform targeted healthcare strategies.- Worked across contracts to leverage existing entity extraction models, adapting them with clinician input to identify the most relevant terms for Long COVID analysis from medical notes.- Upgraded computational environments to support advanced analytics, testing Azure Machine Learning and Databricks integration to lay the groundwork for sophisticated AI/ML solutions.

Overlanding, catching up with friends, studying LLM's and AI, Building computer networks.

SAIC acquire of Halfaker-Prototype daily runtime automation of NLP models into production for Text analytics of Mental health determine factors and cancer care tracking-Leveraged a secure Docker container provided by Microsoft, adding essential packages to create a robust foundation for a Spark-Standalone cluster. Developed a custom Python loop wrapper to efficiently run client's existing code in a multithreaded manner, making full use of all available cores without requiring code modifications. The resulting system was a milestone in data analysis, as it processes ALL ~1.5M VA free text notes daily (~250K on weekends), scanning for Social Determinants of Health (SDoH) terms to aid in suicide detection. Careful choice of hardware specifications, tailored to the program, maximized operational efficiency by fully utilizing the available RAM, disk space, and processing cores. This strategic approach, coupled with effective cost management in networking, storage, and computing, led to a remarkably low operating cost of $19 per day.

– Performed training in MaskRCNN-benchmark using 18GB / 118k images with 48% the speed of a Nvidia DGX-1 box at 22% of the price on custom cluster – Competed in Kaggle Steel Defect Detection (Computer Vision) competition with a solo PyTorch program incorporating transfer learning for the base classifier. Achieved 86.7% accuracy, Top model was 90.9%– Competed in Kaggle Predicting Molecular Properties (prediction) competition with a from scratch PyTorch program to predict the interaction between atoms given molecular structure placing in the top 88%– Developing an AWS Machine learning platform to identify hail damage in drone videos to determine roof replacement cost or replacement for insurance adjusters – Developing in house on secure cluster running Apache Hadoop with Submarine a crypto-market trading program – Using Unity environment to train a Reinforced Deep Learning robotic agent to help people playing VR avoid collision with their tv– Implemented a Redundant small business server backup system preventing data loss and allowing for quick drawing recovery for a client– Performed data recovery on a drive that was thought to have been wiped, saving thousands of dollars in time. Plus implemented new security credentials to keep the mistake from happening again for a client – Complete network overhaul of small business, enabling centralized file system along with backups preventing data loss due to natural disasters or employee mishaps – Performed next day disaster recovery fixing 12 computer towers, recovering data, and setting up office network after a flood on Friday. They were open for business the following Monday

– Helping Veterans with the Azure Cloud

– Adapted Nvidia NeMo library into the Azure ML environment. Enabling a single easy to use library for Natural Language Processing (NLP) simplifying training of United States Department of Veterans Affairs (VA) employees and quickening collaboration across customer task. – Incorporated transfer learning of Nvidia BioMegatron a state-of-the-art (SOTA) language model for biomedical and clinical NLP to achieve 94% recall on Named Entity Recognition (NER) for disease datasets– Practiced Big Data techniques analyzing over 157 million radiology notes in PySpark on distributed Databricks cluster in under 22 min with 208 vCPUs – Worked with Microsoft Azure technicians for a Proof of Concept Multi-node Multi-GPU Nvidia cluster enabling the training of a specialized VA NLP model for NER inferencing on radiology notes. – Worked with top doctors from Stanford and Yale to develop NLP models for Cancer and Mental Health Suicide Detection. Enabling veterans to get the care they need quicker and accurately. – Incorporated transfer learning of Microsoft TuringNLRv4 model, a SOTA language model for NLP sitting at #2 on the General Language Understanding Evaluation (GLUE) and SuperGLUE benchmark leaderboards.– Performed high-level overview and low-level technical demos of the platform to proposed clients, leading to the onboarding of 3 new clients.

– Implemented Amazon SageMaker Neural Topic Modeling (NTM) unsupervised learning algorithm for group name predictions from a 1.37 million record data set. – Developed a new data cleaning strategy coupled with the NTM algorithm boosted prediction accuracies from 40% up to 59.7% – Researched and Developed RoBERTa implemented in Amazon SageMaker. RoBERTa is a Natural Language Processing system developed by Google in TensorFlow (BERT) Improved by Facebook in PyTorch (RoBERTa). With a cutoff filter and specialized data cleaning strategy improved prediction accuracy to 80.3% – Incorporated transfer learning into RoBERTa, saving hundreds of dollars and hours in GPU training time.– Incorporated transfer learning into XLM-R. Demonstrating successful training on the AWS data lake, transferring the model to Azure, and preforming inference without retraining.– Practiced Agile Ci/Cd pipeline configuration, getting machine learning models into scalable production.– Worked with Amazon Step functions and Glue Task to automate the Data intake, SageMaker training, and model prediction pipeline. – Data-Science repository Admin under the Intelligence pillar of PDXC

– Transfer learning using TensorFlow object recognition to analyze drone footage of cellphone towers for mechanical defects – Implementation of Nvidia - Docker with TensorFlow for easy deployment to AWS SageMaker allowing of efficient scaling up of systems – Networked and optimized in house cluster to train neural networks , saving hundreds of dollars in cloud time . Started cluster optimization using Apache Hadoop Submarine as a resource management platform

– Programmed computer control systems to automate the cleaning of sensitive laboratory equipment – Helped multiple teams overcome difficult design problems, leading them to project completion and ensuring budget requirements were met– Self-directed, requiring limited supervisor involvement

– Led a team for a complete design package of a concrete rail tie car. This included assembly drawings, weldment assemblies and sizing, and detailed part drawings. Car had a unique requirement of a rail alongside the deck for an overhead gantry crane that traveled between cars preforming track repair. Rail car is still in-service today– Programmed impact recorders, retrieved data, and performed vibration analysis to provide a detailed report in optimal shipping configuration of multi-million-dollar nacelle units to arrive without damage. – Designed Schnabel train car deck with load capacities reaching upwards of 900,000 pounds– Collaborated with manufacturers on revisions of rail car designs and production drawings to cover efficient use of materials and manpower– Performed final inspection of heavy-duty rail cars ensuring they met OTLR requirements and were safe to travel by rail– Preformed classical analysis and finite element analysis (FEA) in order to improve designs– Developed documentation of problems encountered to help training of new fixture design teams

– Taught labs for Pro/ENGINEER and Creo Elements – Integrated Solidworks into previous graphics course – Teaching assistant for the new Solidworks program Lab at UTA

– Implemented a GIS data base, so jobs entered by the company are accessible with an easy to use interface (Esri ArcGIS)– Drastically reduced cost and optimized the performance of the field crew– Implemented a central file server for streamline file sharing and drawing modification – Worked alongside North Texas Water District as GIS Technician for twenty-thousand-acre lake project– Implemented redundant backup system, preventing any data loss. Tested successfully a few years after implementation.– Created large topographical drawings using AutoCAD