Research
Papers and ongoing studies in AI fairness, language technology, and graphics.
- 012026First author
Aesthetics Is Not a Mirror
An end-to-end audit of Stable Diffusion v1.5 outputs against the LAION-Aesthetics V2 curation step: what an aesthetic filter keeps, and what a model trained on it learns to see.
Across 15 occupations and 3,000 Stable Diffusion v1.5 generations, the gender composition of the curated data predicts the model's outputs (r = 0.873). 200,000-trial Monte Carlo joint null test, FairFace cross-classifier replication, pre-registered held-out falsification. First-author poster, Midwest Graduate Symposium 2026.
- 022026Dr. Aniruddha Ray Lab, UToledo
DP-4DGS: Dynamic Priors for 4D Gaussian Splatting
Working draft. Coupling 4D Gaussian splat scene reconstruction with a physics-informed prior to improve dynamic-scene fidelity under sparse observation.
First-author preliminary results submitted to the Midwest Graduate Symposium 2026. Manuscript in preparation for a graphics and vision venue.
- 032025IEEE QPAIN 2025, published
Bengali Fake News Detection: A Multi-Layered LSTM Ensemble Approach
A multi-layer LSTM ensemble for detecting misinformation in Bengali-language sources, evaluated at the corpus level.
An ensemble of multi-layered LSTMs trained on a curated dataset of verified Bengali news, with weighted prediction merging, reaching 82.43% accuracy.
- 042026arXiv, second author
Context-Aware Dialectal Arabic Machine Translation with Interactive Region and Register Selection
A context-aware, steerable framework for dialectal Arabic machine translation that models regional and sociolinguistic variation explicitly.
A rule-based data augmentation pipeline expands a 3,000-sentence seed corpus to 57,000 parallel sentences across eight regional varieties; an mT5-base model is fine-tuned on lightweight metadata tags for controllable region and register.