software developer · simulation, systems & ML
Computer Science student at Sacramento State (graduating Dec 2026). I like turning hard problems into things you can watch run — evolutionary simulations, parallelized systems, and machine-learning models — mostly in Python and C++, with side trips into Java and lately Rust.
Off the keyboard: movies, books, games, and wandering around outside.
Full-stack Django platform built from scratch with an 8-person team over 8 months for a family law firm. I helped design the 5-app modular architecture and 3-tier role-based access (Guest / Client / Admin), and built shared infrastructure - Stripe & Calendly webhooks, Google OAuth, and a CKEditor admin interface that lets non-technical staff publish rich-text content site-wide without a developer.
Genetic algorithm for the coverage problem: robots with limited energy evolve movement rules to explore as much of an unknown grid as they can before powering down. Over 500-agent generations they converge on a boustrophedon row-sweep at >90% coverage; OpenMP parallelization cut a full run from 138s to 5s (27x). The live demo is a simplified JS port.
Binary classifier over 320K CDC BRFSS survey records (~8.6% positive)
benchmarking KNN, Logistic Regression, Random Forest, XGBoost, LightGBM,
CatBoost, and an MLP. Engineered clinically grounded features (composite
risk score, comorbidity counts) and selected predictors with LightGBM,
fit only on training folds to avoid leakage. Handled class imbalance via
scale_pos_weight rather than SMOTE, evaluating with stratified
5-fold CV on ROC-AUC, F1, and MCC.
Recursive generator that grows unique cherry-blossom trees with randomized branching, death chance, and petal placement. Python + Tkinter. (Demo simplified and written in JS).
Open to opportunities, collaborations, or just chatting about tech.
Shipping a real product for a real client, with a team of 8.
This wasn't a class toy, it was a live site for a family law firm. Our non-technical client had to be able to run it day-to-day, and her clients had to book and pay without a developer in the loop. That constraint drove most of the design.
Letting selection discover a strategy I never coded.
The question: can simple movement rules evolve to cover an unknown grid better than anything I'd hand-code, under a strict energy budget? Yes, they can.
VISITED cell state.
This gave each bot the ability to learn that there was no benefit to visiting a cell twice.
138s → 5s (27x) in testing, eventually running in <1s on the release build.
Doing the evaluation honestly on 320K messy survey records.
Predict heart-disease risk from a big, imbalanced public-health survey, and resist every shortcut that would make the numbers look good but not generalize.
scale_pos_weight instead of SMOTE, reweighting the loss
rather than synthesizing fake patients, which I could defend clinically.A handful of simple rules, a different tree every run.
I wanted to drill recursion in an interesting way, and generating a physical tree analog seemed like a fun way to do it.
A recursive generator grows a unique cherry-blossom tree every run. Randomized branching, death chance, and petal placement.
50 robots per generation evolve over 50 generations via selection & crossover. The maze renders when evolution finishes. Compare a gen-1 bot against the best evolved one.