San Francisco, CAMember of Technical Staff at Datology

Building multimodal data and evaluation systems at Datology.

I build multimodal systems at Datology. The work spans pretraining, VLM evaluation, distributed training, and research infrastructure. Recent projects range from data curation systems and vLLM eval paths to multi-node training and agentic tooling that shortens experimental loops.

Signature artifact

I care most about shrinking the distance between a data decision and a trustworthy model comparison.

Featured research

DatBench

DatBench is the clearest public example of the kind of work I care about: evaluation systems that shape model decisions and still stay inside the loop.

DiscriminativeFaithfulEfficient
DatBench figure
Crop A
DatBench crop
Eval loop
arXiv2026

DatBench: Discriminative, Faithful, and Efficient VLM Evaluations

A benchmark and evaluation framework for vision-language models built around discriminative tasks, faithful scoring, and practical efficiency.

S. Joshi, H. Yin, R. Adiga, R. Monti, A. Carranza, A. Fang, A. Deng, A. Abbas, et al.

Current work

Systems work on the critical path

Three operating lanes keep the loop moving. Together they cover benchmark design, data handling, and launch.

Primary lane
01

Benchmark and evaluation systems for VLM research

Built evaluation paths that keep model comparisons useful for curation decisions and model iteration.

Current lane
02

Multimodal data curation and export pipelines

Built ingestion and export paths that make large multimodal corpora easier to train on and inspect.

Current lane
03

Distributed training and launch infrastructure

Added vLLM eval support and hardened multi-node launch plus checkpoint behavior for faster experimental turnover.

Selected public work

Open work that still points in the same direction as the current systems surface. Fewer pieces, clearer signal.

Benchmark Dataloader

Benchmark Dataloader

A benchmarking setup for multimodal dataloaders, built to surface throughput bottlenecks before they become training-time surprises.

GitHub
SpecReFlow

SpecReFlow

Reflection-aware video restoration research translated into a public implementation for medical imaging workflows.

GitHub