I just wrapped a B.A. at UC Berkeley in Computer Science and Data Science (econ emphasis). Right now: applying for two-year research-assistant programs — empirical finance or computational topology — as the bridge to a PhD, while moonlighting as a senior backend engineer.
The thread that keeps me curious: finding hidden structure in complex data — whether it's a tax-policy effect across Michigan census tracts, a network of capillaries in a retinal scan, or the symmetry of a 2D shape encoded in its Euler Characteristic Transform. I want the math and code that make the structure visible.
Cooking got me here first. As a kid I'd spend hours with a recipe and a YouTube tutorial; what mattered was the process, not the dish. Programming, then research, gave me the same shape — only the medium changed.
Before Cal I grew up in Liberty, Missouri — 9 years as a floor tech for the family carpet-cleaning business, ProStart culinary competitions (3rd at nationals), and a small obsession with making things that work. Off-the-clock: still heavy in the kitchen, captain in the RBLN, and two years as a Berkeley RA.
Writing + code I send with grad-school and predoc applications.
Pitted Transformers, LSTMs, GRUs, and vanilla RNNs against DFA learning tasks to study generalization dynamics. Wrote the trainers, ablations, and a 10K-sample eval framework from scratch.
Estimated how monetary-policy surprises (STMT — the principal-component summary of short-rate news) propagate to Treasury yields, breakeven inflation, and exchange rates. Heteroskedasticity-robust regressions across a cross-country panel with NFA/GDP heterogeneity. The heatmap is the cross-currency correlation structure of the FX-return responses.
One honest note: this is a 24-hour take-home, so the writeup has a few rough edges — a figure label or two doesn't quite match the latest output, and rerunning with slightly different data cuts moves the numbers around. The point of the challenge isn't a polished paper; it's whether you can pick the techniques, apply them cleanly, and interpret what comes out. That's what I'd hand someone reading this for that signal.
Class projects and games. Coding samples ↑ and research ↓ get their own sections.
Civic platform to track representatives and political news. Implemented the Issue Classification feature (17 categories), backed by 60+ RSpec tests at 100% pass.
Generates interactive, parallelized cooking timelines. Web scraping, recipe parsing, history tracking — so you can actually run three pots without losing your mind.
Logistic regression, random forest, gradient boosting, clustering — full sweep on clinical + lifestyle features for diabetes risk.
2D tile-based world exploration engine with randomized, reproducible map generation — rooms, hallways, the works.
The classic. Slide tiles, merge powers of two, try not to lose to a 4×4 grid.
Plants-vs-Zombies, but ants. CS 61A's final boss — inheritance, polymorphism, and tower-defense logic. Framework + sprites by course staff; the action methods are mine.
Diploma, four summers of research posters, and the receipts.
A topology-driven framework for detecting bilateral and rotational symmetry in 2D shapes. Compared each shape's ECT matrix to its rotated and reflected versions using L1 distance to find the angles or axes of highest similarity. Synthetic shape datasets (rose curves, Gielis' superformula), polar plots, Fourier spectral decomposition. Taught myself the foundations of topological data analysis along the way — and the work sparked the interest now driving my graduate apps.
With Dr. Elizabeth Munch and Dr. Nkechi Nnadi · SROP
Returned to the project with ACS data through '22, added demographic controls, tested lagged effects, and ran robustness checks across alternative DiD specifications. The analysis showed meaningful drops in poverty and unemployment alongside increases in income and labor-force participation in designated tracts. Presented to academics and policymakers — oral talk at the Midwest Economics Association conference.
With Dr. Jeffrey Wooldridge · SROP
Built a U-Net convolutional neural network in TensorFlow + Keras to segment retinal blood vessels from a novel fundus dataset. Hand-annotated training masks (no existing labels), tuned data-augmentation pipelines, MATLAB preprocessing. Reported every two weeks to ophthalmology residents and the Dean of Ophthalmology. The model's vessel segmentations — capillaries and vascular branches critical for diagnosing diabetic retinopathy — became the backbone of an NIH-funded automated-screening study; drafted a first-author manuscript with my advisor.
With Dr. Yanhui Ma · only undergrad / CS student on the team · SROP
The first summer. Built the original tract-level longitudinal panel of Michigan communities (ACS '12–'22) and independently implemented difference-in-differences and fixed-effects estimators in R and STATA to evaluate the 2017 TCJA Opportunity Zone designation. The foundation the 2024 round expanded — fun to see how the poster design grew up between the two.
With Dr. Jeffrey Wooldridge · SROP
Best way to reach me is email — I read everything, even the cold ones.
toba4366@berkeley.eduCS 61B Project 0 — game framework by course staff (P. N. Hilfinger). The Model · Board · Tile · Side implementation is Trenton's, compiled to a 7 KB jar that runs in your browser via CheerpJ (WebAssembly). GUI is a fresh JS canvas re-render — no staff-licensed library ships.
CS 61B Project 3 — project framework + tile system by course staff. The world-generation algorithm (rooms, hallways, apple placement) is by Trenton with project partner Kamoni Fletcher. Compiled to a 20 KB jar via CheerpJ. The avatar, chaser shapes, and four pathfinding algorithms (Greedy / DFS / BFS / A*) are JS-side. Music: bulletsong.wav.
ants.py: ant action methods, throwing/merging rules, container-ant logic, QueenAnt — is Trenton's. Source is shipped as marshalled bytecode rather than as raw .py to avoid leaking the full course solution. CS 61A ↗
Click an ant card to select, then click a tunnel tile to deploy. Bees enter from the hive (right) and march toward the home base (left). All ants.py classes — Ant, Bee, GameState, QueenAnt, etc. — are running as actual Python via Pyodide (WebAssembly).