projects

Research projects and interactive AI/ML systems — from brain-imaging work to browser demos.

A few things I’ve built. Featured research projects are pieces of my own work on brain imaging and neurodegenerative disease, each with a paper and code. Below them, AI/ML experiments are interactive systems that run live, entirely in your browser. Click any link to open. (Demos are best in desktop Chrome or Edge.)

HiLoGraph framework overview: multi-scale brain graph construction and model architecture MICCAI 2026 · Early accept · top 9%
Research Project · first author
HiLoGraph — Hierarchical-Longitudinal Brain Network Representation Learning
HiLoGraph encodes brain networks at two scales — fine-grained 3-hinge gyral landmarks and atlas regions, built from cortical morphology and structural connectivity — into one shared latent space, with longitudinal constraints across each subject's repeated scans. Trained only on healthy adults, this space serves as a normative reference of brain aging: a patient's deviation from it separates cognitively normal, Alzheimer's disease and Lewy body dementia without any disease-specific fine-tuning.
Paper · Code
Continuous staging tree placing CN, MCI, AD and LBD along one continuum MICCAI 2025 · Early accept · top 9%
Research Project · first author
A Unified Continuous Staging Framework for Alzheimer's Disease & Lewy Body Dementia
AD and LBD are usually staged with coarse, discrete clinical labels and studied in isolation — but patients lie on a continuum from mild cognitive impairment to dementia. This work places both diseases on a single continuous staging axis built from hierarchical anatomical features, so a patient's position reflects where they truly sit along that continuum.
Paper · Code
Disease Embedding Tree method overview: 3HG features, MMSE-guided embedding and the disease tree AAIC 2025 · Alzheimer's & Dementia
Research Project · first author
Disease Embedding Tree — a Finer-Scale Cortical Representation of the Impairment Continuum
Whole-region summaries are too coarse to capture where cortical disease signatures emerge. The Disease Embedding Tree is a finer-scale cortical representation that embeds the AD/LBD cognitive-impairment continuum at a sub-regional scale.
Disease Progression Explorer
Interactive Demo
Disease Progression Explorer
Slide along the cognitive-impairment continuum, from normal through MCI to dementia, and watch atrophy spread across a live brain network — reddening the medial temporal lobe for Alzheimer's, the occipital cortex for Lewy body dementia — on one shared staging axis. An interactive illustration of my Unified Continuous Staging work.
Try it live →
Brain Connectome Explorer
Interactive Demo
Brain Connectome Explorer
The brain as an interactive network — the seven functional subnetworks, hub regions, and the circuits that break down first in Alzheimer's. A general illustration of the kind of network representation my research learns from brain MRI. Raise the threshold, isolate a subnetwork, or highlight the AD-implicated regions.
Try it live →

AI/ML experiments

Interactive machine-learning systems I built to run entirely client-side.

Browser LLM Chat
Interactive Demo
Browser LLM Chat
A ChatGPT-style assistant running 100% in your browser via WebGPU — no server, no API key.
LLM Agent Visualizer
Interactive Demo
LLM Agent Visualizer
Watch an LLM agent reason step by step — the Reason → Act → Observe (ReAct) loop, animated.
Neural Net Playground
Interactive Demo
Neural Net Playground
Train a neural network from scratch (backprop hand-written in JS) and watch it learn.
Attention Explorer
Interactive Demo
Attention Explorer
See what a Transformer looks at — a real BERT's self-attention across every layer and head, live in your browser.
Semantic Search
Interactive Demo
Semantic Search / RAG
Search by meaning, not keywords — real sentence embeddings in your browser (the core of RAG).

More experiments

Smaller demos: Tokenizer Playground (the real GPT-4 BPE, live) · LLM Knowledge Graph (any topic as a force-directed graph).

Demos are built with vanilla JavaScript; the models run client-side via WebGPU / WebAssembly. The text you enter never leaves your browser — only the model weights are fetched once from a public CDN.