hi, i'm Bumjin

I'm a Machine Learning Engineer working towards an MSc in ML/AI @ Brown CS. I love learning about how machines... learn! While I'm not engineering, you can usually find me in the dance studio with my team.

Github

Resume

experience

Radiology AI Lab - RI Hospital

Student Researcher

Providence, RIOct 2025 - Present

Currently exploring multimodal modeling techniques to predict pancreatic cancer patient survival outcomes and reaction to treatments. Coordinating closely with medical faculty to incorporate visual pancreatomy slides, textual physicians' notes, and genomic mutation features.

PyTorch

Hugging Face

Python

Git

Slurm

Brown CS

Course Dev, Teaching Assistant

Providence, RIMay 2022 - Present

Graduate Deep Learning

(Sept 2025 - )

utilized GitHub Actions to automatically verify internal code functionality on every update
helped students with transformers, autoencoders, diffusion, and more!

Computer Vision

(Jan 2025 - May 2025)

Extensively unit tested and play-tested assignments to improve assignment quality for users
Hosted live debugging and Q&A sessions for conceptual / coding questions during weekly office hours

Undergrad + Graduate Deep Learning

(May 2022 - May 2023, Jan 2024 - Dec 2024)

developed and refined course content (informational documents, labs, coding assignments) for 300+ concurrent students
prototyped new ViT assignment which used transformers for generative image captioning
created and introduced Beras: a manual reimplementation of Tensorflow & PyTorch auto-differentiation tools for deep, gradient optimization models
helped students on topics including language modeling, transformers, GANs, and more
mentored student groups through research-style deep learning projects

Tensorflow

PyTorch

Keras

NumPy

Jupyter

Colab

Pandas

Git

Bash

Singh Lab @ Brown

Student Researcher

Providence, RIJuly 2022 - July 2025

Broadly explored deep learning models for medical analytics, building code to train and optimize models on Brown's High Performance Computing Cluster. Previously contributed to developing prototypes of multimodal contrastive transformers. Developed my Senior Honors Thesis: GraphSC, An Exploration into and Analysis of Graph Networks in Stem Cell Colony Characterization.

PyTorch

Jupyter

Colab

Optuna

Git

matplotlib

Slurm

Bash

Docker

Mahmood Lab

Student Researcher

Cambridge, MAJuly 2022 - July 2025

Developed Multiple Instance Learning models to make patient survival predictions on gigapixel biopsy images.

See paper here

. Additionally trained GAN models to create high resolution biopsy samples to serve as artificial training data for rare cancers

PyTorch

Jupyter

pandas

sklearn

matplotlib

projects

GraphSC

Report

Github Link

A mutlitaksing graph-based approach to automated stem cell characterization from microscope imaging with a reduction in average squared error of over 50% in each of 3 target variables when compared to previous state of the art models and modern ViT & Foundation Modeling-based approaches. The training time for the vision encoder was reduced from days to just hours. The graph network further achieved rich embeddings, enabling transfer learning onto novel targets with just a linear probe.

PyTorch

Weights & Biases

Jupyter

Pandas

matplotlib

Seaborn

Bash

Slurm

YogaSplat

Report

Github Link

A fast, action-conditioned novel-view generator for pre-existing videos, exploring extensions of 4D Gaussian Splats (4DGS). Designed to generalize 4D Gaussians beyond a single reference video, YogaSplat demonstrates the capability for 4DGS models to encode multiple actions for a single target subject without the need to retrain distinct models.

PyTorch

NumPy

Jupyter

Pandas

matplotlib

Bash

Slurm

MAE Fine Tuning

Github Link

An exploration into the Self Supervised Learning properties of the ViT-based

Masked Autoencoder

. This project partially reused model code, but also relied on manual rebuilds of the training regimes and Distributed Data Parallel to ensure compatibility with Brown Oscar CCV while reducing training time by over 30%. ViTMAEs were pretrained to reconstruct unlabled, masked images and fine tuned by linear probing the encoder's embeddings on a classification task. Ablation studies into optimization techniques like using mean pooling encodings instead of classification encodings, linear probe intitialization, choice of optimizer, learning rate decay method, and batch normalization demonstrated the relevance or lack of impact from such techniques in achieving optimal performance.

Pytorch

Torch Distributed

Weights & Biases

Hugging Face

Slurm

ImTex

Report

Google Drive Link

A from-scratch language model for converting images of math formulas to their associated LaTeX code. Given a screenshot of a formula, the image encoder - transformer decoder model predicts out a LaTeX formula. Certainly not perfect, but still decoded LaTeX-parsable code quite close to ground truth.

Python

Colab

PyTorch

NumPy

matplotlib

BertQA

Google Drive Link

A repurposed autoregressive language model trained to highlight answers to questions in a body of text. Achieved 89% precision, recall, and f1 score in a span of hours when fine-tuning a general next-token BERT model on the

SQuAD 2.0 dataset

Python

Google Cloud Platform

PyTorch

Hugging Face

NumPy

NLTK

community involvement

DAEBAK Dance Team

Google Drive Sample

Vice Director

(Sept 2024 - May 2025)

developed a real-time system of Google Sheets for our Performer Casting List, Performer Practice Availability Aggregation, and Audience Member Ticket Aggregation
- efficiently and automaticallly aggregated information through spreadsheet manipulation techniques (Sheets Custom Functions, SQL, RegEx)
- A/B tested forms to encourage faster response times from members
- visualized aggregate availability with custom filter rules into easily understood calendars
- reduced leadership workload by weeks
Coordinated member availability to schedule practice and performance spaces from Brown University campus, averaging 8 hours of booked practice spaces per week
Managed over 100 different performers in 30 unique pieces
Flexibly adapted to diverse tasks as needed (Figma graphic design, social media management, performance media editing)

Performer, Teacher

(Sept 2021 - May 2025)

Hosted dance workshops to teach community members selected pieces of choreography
Performed at annual show case with over 800 audience members over 2 nights.
Check it out here!
Designed stage lighting and animations for select pieces

Google Sheets

SQL

RegEx

Google Forms

Figma