Work
Brevity is the Soul of Inference Efficiency
Response length, not just model size, as a lever for inference cost in VLMs. Curated data that keeps models concise cuts inference cost up to 35x while holding accuracy within a point.
arXiv, 2026. DatologyAI.
20/20 Vision Language Models
Data curation alone, no architecture or training changes, closing most of the gap to frontier VLMs. A 2B model beat InternVL3.5-2B by 9.9 points at 17x less training compute.
arXiv, 2026. DatologyAI.
The Finetuner's Fallacy
Finetuning isn't always the cheapest way to specialize a model. Folding domain data into pretraining instead cuts the tokens needed to reach a given domain performance by up to 1.75x.
arXiv, 2026. DatologyAI.
ÜberWeb
Multilingual pretraining data across 13 languages, showing that most "multilingual interference" is a data quality problem, not a fundamental limitation. Released a 20-trillion-token public dataset.
arXiv, 2026. DatologyAI.
DatBench
A cleaner suite of VLM evals: up to 70% of some existing benchmark questions turned out to be answerable without looking at the image. Filtering fixed that and got a 13x average speedup.
arXiv, 2026. DatologyAI.
Luxical
Open-source library for fast text embeddings, blending sparse lexical features with a small distilled model. 3–100x faster than neural baselines at competitive quality.
arXiv, 2025. DatologyAI.
BeyondWeb
A framework for synthetic pretraining data at trillion-token scale. Smaller models trained on it match or beat larger models trained on competing synthetic data.
arXiv, 2025. DatologyAI.
Experiment Harness
Automated orchestration that compressed month-long research cycles to weekends. Built because I was tired of babysitting runs.
Datology.
VLM Eval Stack
Migrated VLM evaluation from HuggingFace to vLLM-based inference. 10x faster eval cycles. Built because the gap between running an experiment and knowing whether it worked was too long.
Datology.
VLM Data Curation Pipeline
Curation at multi-billion sample scale for CLIP pretraining. Curated data matched uncurated performance 10x faster, 2x faster than CLIPScore filtering.
Datology.
UniCat
Showed that training sensor modalities independently and concatenating at inference beats joint fusion for multimodal re-identification. The result that made me take data-level decisions seriously as a research variable.
NeurIPS 2023 (UniReps Workshop). Modern Intelligence.
GraFT
Gradual fusion transformer for multimodal re-identification. Fusion tokens inside a ViT backbone. 62% smaller than comparable models, state-of-the-art on several benchmarks. The predecessor to UniCat.
Modern Intelligence.
Digital Staining for Bladder Cancer Detection
Extends the SpecReFlow line of work: label-free digital staining that converts white light cystoscopy video to blue-light-equivalent, without the dye.
Communications Medicine, 2024. Bowden Lab, Vanderbilt.
SpecReFlow
End-to-end pipeline for detecting and restoring specular reflections in cystoscopy video using optical flow. My first published paper, and the project that taught me where the real problems in perception live.
Journal of Medical Imaging, 2024. Oral presentation, SPIE Photonics West 2023. Bowden Lab, Vanderbilt.