About Me
I'm Ty Tracey, a Staff AI Engineer at Propelus and, increasingly, a researcher. I spent 13+ years building performance-critical systems; today I work on machine learning for medical licensing and fraud detection, and I'm pointing the rest of my career toward research in AI for science.
What I Do Now
At Propelus I apply machine learning to medical licensing, verification, and fraud detection: retrieval, entity resolution, and data matching over structured, sparse, and noisy data. I built a domain-ontology induction pipeline (TNT-LLM-style, on LangGraph), and I fine-tune embedding models for a two-stage retrieval setup, a bi-encoder for recall followed by a cross-encoder reranker for precision.
Alongside the day job I follow a few threads: embedding architectures and the tradeoffs in retrieval systems, GPU work down at the kernel level (including a from-scratch CUDA inference engine), and an interpretability direction I've scoped but not yet built, using sparse autoencoders to make ontology induction reproducible. The throughline is moving toward research in AI for science.
Technical Expertise
Machine Learning: PyTorch, transformers and representation learning, two-stage retrieval, embedding models, inference optimization, and a working grounding in mechanistic interpretability (sparse autoencoders).
GPU & Systems: CUDA and kernel authoring, TensorRT-LLM, Triton, OpenGL/GLSL, performance work, and distributed systems.
Languages: Python, C/C++, CUDA, Rust, TypeScript, Go, Java, and Clojure.
Engineering: CI/CD, testing, cloud, and full-stack development.
Background & Mission
BS in Computer & Information Systems from the University of North Florida, then 13+ years building performance-critical systems: 3+ years at Meta on an internal VS Code IDE and dev infrastructure, plus stints at Cisco/Duo, VERB, SemanticBits (healthcare/HIPAA), and earlier enterprise roles.
The trajectory: This isn't a pivot so much as a deliberate move from building systems to doing research. I'm self-studying geometric deep learning, proof-based math, and the physics-to-chemistry-to-biology stack, and going deep on GPUs so I understand models at the kernel level. The destination is research in AI for science.
Outside Work
I'm a guitarist and built FourthsHub, an all-fourths guitar learning platform, and I still dabble in graphics and gamedev; these are hobbies now, not the focus.