Junseok Oh

Hi! I'm Junseok, pursuing an MSE in Data Science at the University of Pennsylvania. My goal is to develop human-centric social robots that support caregivers and patients through multimodal interaction.

I study multimodal social attention: how an agent can infer who needs attention, what matters in the moment, and when assistance is appropriate from cues like gaze, head orientation, language, and turn-taking. At the MIT Media Lab Personal Robots Group (mentored by Dong Won Lee and Dr. Hae Won Park), I work on multimodal egocentric head-gaze prediction. Previously, I worked on health-focused projects spanning smartphone-based biosensing and single-cell analysis. I also have industry experience at Hewlett Packard Enterprise (HPE), where I worked full-time as a Cloud AI Software Engineer and previously interned in Data Science.

I earned my B.S. in Computer Science from Purdue University, with a minor in Mathematics.

Email  /  CV  /  Scholar  /  GitHub

Research

We study social egocentric head-gaze prediction by fusing vision embeddings with speaker-aware audio and language cues.

We develop a smartphone-integrated, optomechanical dual-mode biosensing instrument that combines vision-based readout with frequency-shift measurements to enable low-cost, rapid detection of Salmonella Typhimurium for practical food-safety monitoring.

Research Projects

K-BioX - Stanford Cardiovascular Institute
Jan 2022 – Aug 2022 and Jan 2024 – May 2024
Advisors: Dr. Siyeon Rhee and Prof. Joseph C. Wu

Built an automated scRNA-seq pipeline over public GEO datasets to compare adult vs. P12 developmental stages, then used GSEA, cell-to-cell interaction profiling, and PCA/UMAP to surface adipogenesis/angiogenesis signals across epicardial adipose tissue (EAT) datasets.

Wittgen Biotechnology - UC Berkeley SkyDeck
May 2022 – Aug 2022
Research Intern

Worked on ML for high-resolution tumor classification and tailored drug recommendations by processing RNA-seq data from 40+ cancer patients, contributing to an AI platform for cancer heterogeneity profiling, and using Seurat for single-cell analysis and cell-type differentiation.

GitHub Projects

A lightweight repository for real-time object recognition experiments and evaluation.

Data collection and preprocessing utilities for egocentric video datasets.

Modeling and experimentation code for egocentric perception tasks.

Teaching

Template based on Jon Barron's website; see Leonid Keselman's Jekyll fork.