Projects

/projects

Product development, machine learning, research, and web development.

This page highlights a selection of projects across product development, machine learning, research, and web development. Some are polished applications, while others reflect experiments, studies, and technical exploration across different areas of my work.

ProductResearchFull-stack

In progress · M.S. Thesis

Teaching models to understand human activity

My M.S. thesis on teaching models to understand everyday human activity from multimodal biosignals. It is still training, details soon.

## Dataset

The work uses the EASE Table Setting Dataset (Meier et al., Scientific Data, 2026): 78 sessions of people setting a table, each recorded with eight synchronized streams, including motion capture, first-person and room video, eye tracking, EMG, EDA, acceleration, audio, and EEG. It is annotated by phase, activity, motion, and the objects people interact with.

PythonMultimodal biosignalsRepresentation learning

Dataset

Aug. 2026

Internal Tool / UAA Dean’s Office

Certificate Automation System

An internal automation tool for the University of Alaska Anchorage Dean’s Office that reduced a repetitive certificate-generation workflow from approximately 3 hours of manual work to about 5 minutes.

per batch: 3h → 5min
less processing time: ~97%
faster workflow: ~36×
certificates per batch: 100+

## The problem

I was tasked with generating individual certificates for more than 100 students. The existing process was almost entirely manual: take a template, type each student’s name, export a PDF, name the file, and repeat for every recipient. A single batch took about three hours.

## The solution

Staff now drop the student roster and certificate template into a Google Drive folder and hand that folder to the app. It verifies the files, reads and normalizes names, finds the name placeholder in the template, generates a PDF for every student, and organizes the results into an output folder. It is built for non-technical staff: a few clicks, file validation, progress, clear error messages, retries, and the ability to stop and resume a batch.

## Impact

The tool is now used by the Dean’s Office for certificate generation, freeing staff time for higher-value work.

Google Apps ScriptJavaScriptHTML/CSSGoogle Drive API

2022 – now

Product / Full-stack

PomoFriends

PomoFriends is a free, ad-free Pomodoro timer built around accountability. You can focus solo, or open a focus room and work alongside friends, "study with me" style. I first built it in 2022 and rebuilt it from scratch in 2026 with SvelteKit.

room modes, sync & async: 2
releases since the Aug 2026 revamp: 5
ads, free to use: 0
current version: v3.2

## The problem

There are plenty of Pomodoro apps, but almost none of them are social. The few that support groups have no shared timer, no chat, and no way to see who else is working, so nothing keeps you accountable.

## The solution

Focus rooms come in two modes. Sync rooms share one timer that the whole group follows; async rooms let everyone run their own timer while still seeing who is focusing. Chat opens during breaks and closes during focus, so it never becomes the distraction. Built-in tasks let you estimate and track pomodoros per task.

## Today

The 2026 revamp put rooms at the center of the app and added stats, streaks that follow your timezone, a leaderboard, themes, and guided tours. The timer, tasks, and stats work without an account; signing up adds rooms and syncs your data across devices. Product usage is measured with PostHog.

TypeScriptSvelteKitFirebasePostHog

Live

Sep. 2022

Web Platform / Full-stack

sup4bubb4

A web platform for the VTuber community that compiles song requests from Amelia Watson’s superchat-reading streams and syncs them with the recorded VODs on YouTube.

## The problem

During these streams, song requests come in through bubb4bot (Twitch channel). Once the stream ends, the request queue is gone, and anyone watching the VOD later has no way to see which songs were requested and when.

## The solution

A Firebase Cloud Function checks for new livestreams and triggers collection of the request queue from the StreamElements API. The site pairs each VOD with its queue: pick a VOD, and the embedded YouTube player and the bubb4bot queue sync up for you.

## Details

Fully automated after setup, with no manual work per stream. The interface is responsive, so it works on phones as well as desktops.

TypeScriptSvelteFirebaseCloud FunctionsTailwindStreamElements APIYouTube embed

Live GitHub

Dec. 2022

Research / Study Design

Predicting Student Learning Outcomes using EEG Correlation with Teacher/Expert Neural Activity

A research study design exploring whether a student’s learning outcomes can be predicted from how closely their brain activity tracks the teacher’s and experts’, measured with wireless EEG in a naturalistic setting.

students, planned: 100
lecture conditions: 3
teacher + expert recordings: 1 + 2
core measure: ISC

## The question

An fMRI study (Nguyen et al., 2022) showed that learning is reflected in brain-to-brain correlation between teachers and students. It was still open whether the same holds with wireless EEG outside a scanner, and whether that correlation can predict a student’s learning outcomes.

## Study design

Record the teacher’s EEG while giving a lecture and two experts’ EEG while watching it. 100 students watch either the original lecture, a shortened version, or a scrambled version, with tests before and after. Inter-subject correlation (ISC) between each student and the teacher and experts is then compared with their learning gains.

## Why EEG

EEG has lower spatial resolution than fMRI, but millisecond timing, low cost, and portability mean it can run in a real classroom. If the correlation predicts learning, it could help evaluate teaching strategies and personalize learning plans.

EEGInter-subject correlationNeuroscienceStudy design

Read