Mathematical AI Researcher · Ph.D. Candidate · University of Central Florida
I work on the mathematics underlying artificial intelligence — tensor decomposition, low-rank structure, random matrix theory and numerical linear algebra — to understand how large transformer models represent learned information and to make them more efficient to adapt and compress.
Hover to rebuild the mesh
6
First-authored papers
accepted in 2026
840×
Fewer adaptation parameters
than LoRA on LLaMA3-8B
4.76/5
Instructor effectiveness,
Spring 2026 (dept. 4.21)
2.01M
Minute-level observations
in the forecasting study
Research focus
I study the mathematics underlying artificial intelligence — tensor decomposition, low-rank structure, random matrix theory, and numerical linear algebra — to understand how large transformer models organize learned information and why they can often be adapted in surprisingly low-dimensional spaces.
I use this structure to develop methods for adapting, compressing, and understanding large AI models with far fewer trainable parameters and lower computational cost. A central focus is CRAFT, a parameter-efficient fine-tuning framework that exploits shared structure across transformer layers, with the broader goal of making advanced AI more efficient, affordable, and accessible.
I also develop systematic trading models and equity-market simulations using machine learning, time-series modeling, and quantitative methods.
𝒳 ≈ 𝒢 ×₁U⁽¹⁾ ×₂U⁽²⁾ ×₃U⁽³⁾
Direction 01 · Adaptation
Decompose the pre-trained weights, freeze the factors, train only what is small.
Direction 02 · Structure
Spectral and statistical analysis of pre-trained projections across 11 models.
Direction 03 · Markets
Frozen time-series foundation models, corrected by small neural and classical heads.
Headline result · COLM 2026
LoRA-CRAFT applies Tucker decomposition to attention weights stacked across layers, freezes every resulting factor, and trains only small square transformations on those factors. On LLaMA3-8B it beats the reported LoRA commonsense average while training roughly one eight-hundred-fortieth as many adaptation parameters.
Selected papers
An extremely parameter-efficient fine-tuning method that performs full Tucker decomposition by higher-order SVD directly on pre-trained weights organised as cross-layer 3D tensors, freezes all resulting factors, and adapts the model through lightweight trainable transformations on each factor matrix.
A comprehensive ablation on high-frequency stock prediction: a frozen 200M-parameter TimesFM backbone paired with small learned correctors, evaluated over 2,011,399 one-minute observations across ten technology stocks.
Marchenko–Pastur random matrix theory separates each attention projection into a random-like bulk and a set of spectral outliers. Zeroing the outliers in Mistral-7B drives HellaSwag, MMLU and PIQA close to chance; zeroing a count-matched bulk subset does far less damage.
Talks & scholarly presentations
Conference presentations of my work on efficient foundation models, tensor methods, transformer structure and model compression.
A presentation on correcting a frozen 200M-parameter time-series foundation model with lightweight neural and classical components for high-frequency financial forecasting.
A tensor-decomposition study of the structure and compressibility that emerges in parameter-efficiently fine-tuned transformer attention weights.
Expected scholarly presentation of CRAFT, a cross-layer Tucker/HOSVD approach to extremely parameter-efficient adaptation of transformer attention.
Expected presentations of accepted ICMLA 2026 work on magnitude-profile attention-head pruning, component-level quantization sensitivity, and spectral outliers in transformer attention.
Teaching
Calculus II, Calculus III and Linear Algebra at UCF, and a full service load at Sam Houston State before that. Student evaluations have stayed between 4.6 and 5.0 out of 5.0, consistently above both department and college averages.
Made clear efforts to engage students4.88
Helpful in responding to questions4.81
Enhanced my understanding of the material4.69
Education
Graduate training spanning mathematical foundations, computer vision, and quantitative finance.
Expected Fall 2026
University of Central Florida · GPA 3.7
Advisor: Prof. Marianna Pensky
2024
University of Central Florida · GPA 3.8
UCF reported its computer vision research as No. 8 nationally in 2025.
2017
Sam Houston State University
2014
University of Colombo
2012
University of Colombo
Contact