Geometry-aware learning
Graph and kernel methods that adapt to local structure instead of assuming one fixed scale works everywhere.
Ph.D. Candidate · University of Southern California
Machine learning, signal processing, and geometry-driven methods for real-world data.
I build methods for kernels, graphs, representation learning, and scientific AI, with a focus on problems where structure matters and generic models are not enough.
Graph learning
Kernel methods
Medical imaging
LLM embeddings
Research
What I work on
Representation quality
Studying whether learned features actually encode useful geometry, stability, and interpretability.
Scientific and medical AI
Working on signal, image, and data-analysis problems where robustness matters more than leaderboard theater.
LLM embeddings in structured tasks
Using language embeddings where semantics can help organize, fuse, or label multi-source data.
Projects
Featured work
Paper + code
Geometry-Aware Graph Construction
A geometry-driven graph and kernel construction method that adapts spectral bandwidth to local structure, improving downstream inference when one global scale is too crude.
Project
GeomFuse
A label-fusion pipeline that uses LLM embeddings to organize geometric context and improve multi-label completion under weak supervision.
Signal
Non-negative Kernel Graphs
A line of work on kernel graph construction for time-series and signal data, including geometry-aware interpretations for neural representations and EEG-style analysis.
Publications
Searchable publication cards
MLSP
Geometry-Aware Graph Construction via Adaptive Spectral Bandwidth Control
We propose a per-node bandwidth criterion for Gaussian-kernel graph construction that matches the effective rank of the kernel to local intrinsic dimension, improving graph-based inference across multiple self-supervised encoders.
Asilomar
GeomFuse: Geometry-Driven Label Fusion using LLM Embeddings
A geometry-aware label fusion method that combines structured embeddings and language-based priors to strengthen weakly supervised completion and fusion tasks.
Signal
Non-negative Kernel Graphs for Time-Series Signals
A framework for representing signals with sparse non-negative kernel graphs, extended to incorporate explicit delays for time-series analysis and EEG-style data.
Theory
Geometric Interpretation of Deep Features
An analysis of neural feature geometry using kernel graphs and polytope structures to explain when representation learning is stable, useful, or misleading.
Blog
Markdown-friendly posts
Why bandwidth matters more than people admit
A short post on why kernel scale choice changes geometry, rank, and downstream behavior.
Read postFrom paper to website
How to turn one project into a paper page, blog post, talk slide, and GitHub link without making a mess.
Read postWhen LLM embeddings help structured data
A practical look at where language embeddings actually contribute signal in non-text problems.
Read postMedia
Embedded videos and podcast links
Conference talk
Embed a YouTube recording of a talk, seminar, or presentation here.
Podcast / demo
Use this space for a podcast RSS player, a demo reel, or a public interview clip.
Background
Experience and education
2019 - present
Ph.D. Candidate, USC
Electrical Engineering, research in graph learning, signal processing, and geometric methods.
2018 - 2019
Research Engineer, Huawei
Video understanding, ad placement, censorship, and computer vision systems.
2015 - 2018
M.S., Bilkent University
Electrical and Electronics Engineering, with research on Magnetic Particle Imaging.
Contact
Let’s keep it simple
For collaborations, talks, or project questions, use the links below.