CVThomas Hayama

Machine Learning Engineer · Sunnyvale, CA

thomashayama@gmail.com · thomashayama.com · github.com/thomashayama · linkedin.com/in/thomas-hayama

Download PDF Open PDF

Applied AI engineer with 3+ years of experience developing production and prototype systems. Experienced with integrating large language models into intelligent tools and deploying scalable services for LLM-based applications. Strong theoretical foundation backed by practical research experience.

Experience

Machine Learning Engineer

The Aerospace Corporation · Data Science and AI Department

– present

  • RAG-enabled chatbot with table understanding. Created a custom document-chunking microservice API on Kubernetes, used ColBERT for embedding and retrieval, and Command R / GPT-4o as the frontend LLM.
  • Arbitrary-length document generation. Used a hierarchy-structure generation method to create long-form documents (longer than the context length) with generated LLM prompts.
  • High-frequency anomaly detection. Used a mix of classical and ML techniques to detect out-of-family behavior in live and historical high-frequency rocket data. PyTorch, Kolmogorov–Smirnov, FFT, CNN
  • Visual odometry for non-cooperative RPOD. Created pipelines for simulated images of space objects in Blender. Implemented deep learning and classical algorithms for 6/3/2-DoF pose estimation, for real-time and offline use. ROS 2, OpenCV, ICP, PnP, ORB, SuperPoint, SuperGlue/LightGlue

Technical Intern

The Aerospace Corporation · Data Science and AI Department

–

  • Applied unsupervised and semi-supervised learning methods for high-frequency anomaly detection in rocket data. PyTorch
  • Modified an automotive LiDAR SLAM library (MULLS-SLAM, C++) for use in space, removing assumptions to allow 6-DoF rather than ~3-DoF estimation.

Research Fellow

Autonomous Robotics and Control Lab, Caltech

–

  • Used PyTorch to train a model to classify thermal images of people and vehicles using only visible-wavelength data (blind domain adaptation) to 95.7%. Experimented with preprocessing techniques to simulate thermal data from visible data, and used variational autoencoders to better transform from the visible to the thermal domain with satellite images.

Projects

Puzzle-solving robot arm

Group project · Python, ROS

  • Given the solved image, solves a shuffled puzzle with a custom five-axis arm, vacuum end effector and RealSense depth camera.
  • Used OpenCV for piece, corner and edge detection for position and orientation.
  • Reached millimeter precision with PID tuning and small relative motions so that pieces fit together; Cartesian motion control with inverse kinematics via Newton–Raphson.

SLAM robot

Individual project · Python, ROS, Gazebo

  • Probabilistic roadmap (PRM) and A* navigation on a two-wheeled robot.
  • Localization via ICP with a 2D point cloud from a RealSense depth camera.
  • Voxel-based mapping, with “probability” built up over time from point-cloud ray tracing.

Fermentation tracker and optimizer (AlcOpt)

Individual project · Python

  • Web app for tracking wine and mead fermentations, recording tasting reviews, and optimizing taste and quality. Recommender system using collaborative filtering and Bayesian optimization.
  • Self-hosted Streamlit site with Google OAuth, in a Docker container on a Linux VM on my Proxmox machine.