Jayden Lee
I'm a Data Science student at UC Berkeley, with a domain emphasis in
applied mathematics, who builds machine learning systems for data
that doesn't behave: behavioral logs, mass spectrometry output,
clinical time series.
Most of my work has been the unglamorous half of ML: getting 300,000
files through a pipeline without it falling over, deciding what a
false discovery rate should be, lining up 36 spectral files that
three tools each name differently. I like problems where the modeling
is only interesting once the data engineering is honest.
- Study
- B.A. Data Science, applied mathematics emphasis, UC Berkeley
- Degree
- Expected May 2029
- GPA
- 3.7
- Group
- Sports Analytics Group at Berkeley · Head of Business
- Languages
- Python · R · Java · TypeScript · SQL
- Modeling
- PyTorch · TensorFlow · scikit-learn · NumPy · Pandas · Seaborn · Matplotlib
- Toolchain
- Git · Docker · Kubernetes · Flask · React · Supabase
- Speaks
- English · Korean
1st place · 7th Annual Datathon at Berkeley
Clinical Intervention Simulator
A Monte Carlo framework that models 300 patient deterioration
scenarios, feeding a Random Forest that optimizes interventions in
real time.
Screenshot · 21:9
The simulation is vectorized with NumPy so a full sweep runs in
seconds rather than minutes, which is what made it usable as an
interactive tool instead of a batch job. It ships as a Streamlit app
over a Pandas ETL pipeline that reconciles multimodal health data into
one frame.
Built at the 7th Annual Datathon at Berkeley, where it took first
place against more than 50 teams and over 200 participants.
- Stack
- Python · scikit-learn · NumPy · Pandas · Streamlit · Tableau
- Scale
- 300 patient scenarios
ChessBlitz
An open-source teaching platform in use at Berkeley Chess School.
Screenshot · 16:10
Flask and Supabase behind the storage layer, OpenRouter generating
hints tailored to the position a student is actually stuck on, and a
React dashboard where a teacher runs their own classroom: rosters,
assignments, and who is struggling with what.
- Stack
- Flask · Supabase · OpenRouter · React · TypeScript · SQL
- Status
- Open source, in classroom use
This Website
The machine you are reading this in. A desktop OS in vanilla
TypeScript — no framework, no UI library, no CSS framework — built
as a study in running AI agents on a codebase without losing control
of it.
The build is multi-agent. Each task routes to a purpose-built skill
— design, code review, QA — and agents run in isolated git worktrees
so several can work at once without colliding in a merge. That only
holds if the agents share a source of truth, so the rules live in
machine-readable specs beside the code: what the product is, what the
visual contract allows, what the game may and may not do. An agent
that has read them cannot argue the interface into something else
three sessions later.
Every change is gated behind a TypeScript build and an automated
accessibility pass before it can land, and each release goes into a
Keep a Changelog log under semantic versioning. The result is a
codebase in the thousands of lines that has stayed consistent across
dozens of agent sessions — which is the actual difficulty, and the
reason this is on the list at all.
The wallpaper is a contour plot of a two-minimum loss surface. The
surface function has one definition, imported by both the build
script that writes the static SVG and the canvas that deforms under
your pointer, so the two can never disagree about the shape of the
field.
- Stack
- TypeScript · Vite · Claude Code · Git
- Dependencies
- 2, both build-time
Experience
-
Feb – May 2026
Data Analyst
Cal Athletics · contract
A pipeline through Snowflake and Colab that quantified purchases
and fed a dynamic fan scoring system, plus clustering and
supervised models predicting ticket-purchase propensity from
behavioral data.
Takeaway Understanding the data
fully accelerates the work done after.
-
Sep – Dec 2025
Business Analyst
Oakland Roots Sports Club · contract
Built a fan-engagement scoring system in Python that folded
real-time BART transit data into a marketing dashboard, and did
the exploratory work behind it in Matplotlib and Seaborn.
Takeaway The mentor won't report
to you — you ask, and you report to them.
-
Jun 2024 – May 2025
Modeling & Simulation Intern
Pacific Northwest National Laboratory
A target-decoy statistical framework modeling false discovery
rates for peptide identification. Deployed the GPU-accelerated
proteomics workflows on the Tahoma HPC cluster under Slurm,
configuring CUDA and batch processing so identification
validation scaled across large mass-spectrometry datasets.
Takeaway Do not assume a result
is correct because it arrived. Ask the statistics whether it is
wrong.
-
Jun – Aug 2023
Computational Biology Intern
Pacific Northwest National Laboratory
Built the file mapping and parsing pipeline that aligned
36 MGF spectral files across tool
formats that each named them differently, then pushed
300,000+ mass-spec files through
GPU-accelerated Casanovo inference on TensorFlow.
Takeaway I like GitHub, and I
want to do more open-source work.
B.A. Data Science, applied mathematics emphasis,
UC Berkeley · GPA 3.7 · expected May
2029
Sports Analytics Group at Berkeley — Head of Business
Coursework: Optimization Models in Engineering · Data Structures and Algorithms
· Data Engineering · Multivariable Calculus · Linear Algebra and
Differential Equations
- Kind
- PDF document
- Size
- 185 KB
- Pages
- 1
Download
spots.map
A few blocks of Berkeley, and what I ate on them. Walk it and the
places open one at a time.
Draft. Not finished, and visible so it cannot be forgotten.
Places I actually went
Twelve or so places in one Berkeley neighborhood, and what I ate at each. Walk the map and they open one at a time.
No places written up yet — the map below shows the storefronts standing empty, which is the truth of it.
Every place on this map is somewhere I actually went, written up in
one file I edit by hand. Nothing here is generated and there are no
filler entries — where I have not written anything, the storefront
stands empty and the map says so rather than inventing a restaurant
to fill the gap. That constraint is most of why it is worth walking.
It is a map rather than a list because a list gets skimmed. Twelve
names and twelve one-line reviews are four seconds of scrolling and
nothing retained. Making you walk two blocks to the next one buys
each place the few seconds it takes to actually read it, which is the
only thing I wanted from the format.
There is no score, no timer and no way to lose, and it does not ask
you to come back tomorrow. Walking is one keypress per block, the map
redraws only when you find something, and nothing on this page
animates while you are standing still.