Aaron Acton

About Aaron Acton

Critical-thinking problem solver with demonstrated success in developing innovative and patented solutions to the most challenging of technical problems. Known for leveraging deep learning, natural language processing, and computer vision within cross-functional teams to deliver impactful solutions to complex real-world challenges.

Experience

1. Principal Data Scientist

   Ralliant

   Jul 2023 — Present · Pittsburgh, PA, US

   Deployed Projects• Customer Insights Analysis System: Developed a speech-to-text (whisper) and language model (GPT-4) pipeline to analyze historical customer conversations for near-real time insights across 16 different metrics, enhancing the organization’s understanding of customer interactions. Made use of structured outputs for reliable parsing.• Binary Classier for Process Error Detection: Developed a scikit-learn binary classier to detect process errors using numerical and categorical data, enabling automation of 60% of a previously manual process. Extracted ~ 10 GB of tabular data, normalized vectors, and stored them in parquet format for scalable access.Prototype Projects• Multi-Class Object Detection and Classication: Created a YOLO (PyTorch) model and entire preprocessing pipeline for real-time object detection and classication to assist with new-employee training. Scaled class count to over 500 and training speed up to 8-GPU instance on AWS SageMaker.• Defect Detection for Manufactured Components: Researched and prototyped a U-Net CNN image autoencoder using PlaidML and Keras/TensorFlow for defect detection, including on-the-y image augmentation. Explored various denoising techniques and Siamese architectures, eventually creating a fully pre-trained defect detection model.• Chat-Controlled Oscilloscope (Patent): Developed a system enabling natural language interaction with hardware devices, translating GPT-based commands into device-specic operations, allowing for more intuitive control of equipment.Experimental Projects• PHI Anonymization Study: Evaluated the effectiveness of various large language models (LLMs) from Hugging Face, including ~ 10 local LLMs (e.g. phi2, llama2, mistral, hermes), ChatGPT, and Presidio, in anonymizing personal health information (PHI). Found that LLMs excelled in categories where out-of-the-box (OOTB) solutions showed limitations, providing better overall accuracy in certain complex scenarios.

Education

• Carnegie Mellon University

  MS, Mechanical Engineering

  2009 — 2012

• University of Ottawa

  BASc + BSc, Mechanical Engineering, Computing Technology

  2001 — 2007

Skills

• Composites
• Software Engineering
• Ansys
• Algorithms
• Random Vibration
• C++
• Perl
• Linux
• Matlab
• Software Development
• Python
• Operating Systems
• Numerical Analysis
• Feedback Control Systems
• Heat Transfer
• Fracture Mechanics
• Solid Mechanics
• Ajax
• Ls-Dyna
• Spacecraft
• Git
• Cad
• Cae
• Solidworks
• Software Design
• C
• Programming
• Optimization
• Spaceclaim
• Dynamics
• Data Structures
• Vibration
• Finite Element Analysis
• Javascript
• Kalman Filtering