Particle Assembly · Identity Through Geometry
CS Researcher & ML Engineer probing the geometry of machine learning — from sparse autoencoders to SSL representation theory and graph embeddings
I'm a Computer Science student at the University of Washington's Paul G. Allen School of Computer Science & Engineering, maintaining a 4.0 GPA while doing research at the intersection of geometric deep learning, mechanistic interpretability, and representation theory.
Currently I'm working as an ML Research Engineer at a stealth AI startup, investigating how CNN inductive biases shape SSL representation geometry — diagnosing alignment, uniformity, and dimensional collapse across architectures. This summer I'll be joining Analog Devices as an AI/ML Intern focused on graph-based ML and embedding models for structured circuit data.
I was previously a Research Fellow at Algoverse AI, where I published RT-TopKSAE at ICLR 2026 — adapting the rotation trick to TopK sparse autoencoders, achieving 100% dictionary utilization vs. the baseline's 47%.
Manifold Hypothesis · Relational Bias · Force-Directed Clustering
Skills as a structured latent space — related tools gravitate into clusters, edges encode learned associations. Hover a node to probe its neighborhood.
General Relativity · Geodesics · Spacetime Curvature
Each role warps the career path around it — like mass curves spacetime. The glowing particle follows the geodesic.
AI/ML Intern — Graph ML & Embeddings
Boston, MA
Graph-based ML and embedding models for structured circuit data.
ML Research Engineer
Research Fellow · Remote
Paul G. Allen School of Computer Science & Engineering
Bachelor of Science in Computer Science
4.0 GPA
Dean's ListEquivariance · Numerical Methods · Geometric Engineering
Designed a VAE enforcing equivariance in the latent space to ensure geometric transformations in input frames correspond to predictable latent trajectories, improving temporal consistency in generated video.
Derived and implemented a full numerical linear algebra engine in pure NumPy — Gram-Schmidt QR factorization, power iteration with deflation, and two-sided Jacobi SVD. Validated against scikit-learn with residuals under 10⁻¹⁰.
Equivariance · Symmetry Groups · Invariant Representations
Community and recognition encoded as geometry. Toggle between force-directed graph and icosahedral simplicial complex — same data, different projections.