Ty Tracey
I work on representation learning and retrieval, and I'm pushing toward interpretability and AI for science.
13+ years building performance-critical systems (3+ at Meta). Now a Staff AI Engineer at Propelus, applying ML to messy real-world data while going deep on the systems and theory beneath modern models.
Currently focused on
Representation Learning & Retrieval
Fine-tuned embeddings and two-stage retrieval for sparse, noisy real-world data
Mechanistic Interpretability
Where I'm headed: sparse autoencoders to make a model's learned features legible
GPU & CUDA
From-scratch CUDA inference kernels, to understand how models actually compute
AI for Science
The long game: applying ML to scientific problems, starting with sequence and protein modeling
What I'm working on
Interpretability for ontology induction
Scoping how sparse autoencoders could make taxonomy induction deterministic and legible, instead of LLM guesswork
CUDA transformer inference engine
Hand-written kernels (MatMul, LayerNorm, GELU) with parity tests, benchmarked against torch.compile
Foundations for AI for science
Geometric deep learning, proof-based math, and a physics → chemistry → biology curriculum