Research projects
Graph State Networks (GSN)
Persistent, nodewise selective state-space models for graph representation learning. GSN keeps per-node state that persists across propagation steps, giving a scalable alternative to deep message passing.
Convergence of Asynchronous SGD for PL functions
Theory and experiments on asynchronous / delayed SGD under the Polyak–Łojasiewicz condition, characterizing convergence behavior between the convex and general nonconvex regimes.
Applied & course projects
Text-to-Image Synthesis Optimization (UCLA MS project)
Optimized text-to-image GAN architectures under hardware constraints and integrated CLIP embeddings for stronger text conditioning.
- Achieved FID 9.85 with a 10.6M-parameter generator as a lightweight alternative to larger baselines.
Mini-GPT (TinyStories)
Implemented a small GPT for story generation, including parameter-saving tricks like weight tying.
- Top-3 token accuracy: ~88.6% on TinyStories.
VOC signature clustering (B.E. thesis)
A machine-learning approach for clustering signatures of volatile organic compounds (VOCs) using E-Nose signals — a low-cost alternative to spectrometry.
- Achieved ~91% accuracy and ~0.89 F1 with a hierarchical pipeline.
Want a shorter view? The Home page has a quick list; this page is the detailed version.