Winxon Nguyen

Autonomy & navigation software, from perception to planning, measured end to end.

UCLA, BS Applied Mathematics (Computing) + BS Statistics & Data Science, Class of 2028. U.S. citizen. Seeking Summer 2027 internships in autonomy, navigation, ground and flight software, computer vision, ML and software engineering.

Résumé (PDF) · GitHub · LinkedIn · winxonnguyen44@gmail.com

About

Applied mathematician who builds the thing, then measures it: offline navigation, real-time sensor pipelines, and GPU kernels. Every claim ships with a benchmark, including the ones that came out a tie.

  • Focus: perception, state estimation, planning, GPU and parallel computing
  • Stack: Python, C++20, Eigen, PyTorch, Numba, Metal, SQL, Django, PostgreSQL
  • Also: founded and sold Twin Digital Media at 18
  • Languages: English, Vietnamese, Mandarin

Selected work

VoroNav: offline navigation for autonomous systems (2026–now, flagship)

A robot navigation stack that runs fully offline: SAM 2 segments first-person video, a from-scratch parallel Voronoi engine (Jump Flooding) builds a maximum-clearance roadmap, and the planner keeps the robot away from walls. The engine has NumPy, multithreaded Numba, fused Metal and C++20 backends, all bit-identical to the NumPy reference.

  • C++20 core (pybind11, hand-written thread pool, GoogleTest, ASan/UBSan/TSan in CI): 19.1 ms for a 2048² diagram, 3.4× faster than 10-thread Numba and 17× faster than the first serial port. The fused Metal kernel is still faster on large grids (14.3 ms).
  • Benchmarked against A*, clearance-weighted A* and RRT* on 94 maps (58K planner runs): minimum clearance matched or beat A* in 99.5% of queries (+70% mean clearance) with 1.5× faster queries; narrow passages are its weak spot (2.1× longer paths).
  • On real SAM 2 masks of first-person video: ≤0.1% collisions vs 0.6–1.1% for grid A*. A temporal filter cuts wall-leak collisions from 14% to 4%.

Python, C++20, pybind11, PyTorch/Metal, Numba, SAM 2. Code and results · C++ optimization log

nav-ekf: state estimation for GPS-denied navigation (2026–now, in progress)

The estimation layer of the VoroNav stack: a 15-state error-state Kalman filter fusing a 6-axis IMU with position fixes, derived by hand with every Jacobian checked against finite differences, with a Python reference and a C++20/Eigen port that matches it to 1e-13 in simulation.

  • Statistically consistent in simulation: over a 50-run Monte Carlo with a 30 s GPS dropout, NIS averaged 3.00 (expected 3) and NEES 15.75 sat inside its 95% χ² band; the filter’s 3σ envelope held 96% of runs at the end of the outage.
  • C++20/Eigen: 1.21 µs per IMU propagation and 1.86 µs per GPS update with zero heap allocations; 55 GoogleTests under ASan/UBSan.
  • On five real EuRoC drone flights the untuned filter proved overconfident during outages; diagnosed in-flight IMU noise at 30–40× the datasheet values. Tuning is next.

C++20, Eigen, Python, GoogleTest, pytest. Code · Results

VoroLens: segmentation-driven Voronoi mosaics in the browser (2026–now)

Lead developer, UCLA Math + Code + Art. An interactive app for the Algorithmic Forms exhibition at the Oceanside Museum of Art (August 2027). Click any object and SlimSAM segments it, then mosaics only that object. SlimSAM runs entirely client-side in a Web Worker on WebGPU with a WASM fallback; images never leave the visitor’s device. Author of Chapter 2 of the lab’s open textbook, Voronoi Tessellation and Adaptive Sampling. A live version runs on this site.

Chapter 2 · Lab site

PairRank: ranking by “better or worse?”, tested on simulated users (2026)

A simulation study plus a small app (Django REST Framework, PostgreSQL, Angular) on comparison-based ranking, inspired by Beli: how few better-or-worse questions place a new restaurant in someone’s list when people answer inconsistently. Every user is simulated with known true preferences; no data is collected from people.

  • With 10% wrong answers in a list of 49, Bayesian bisection cuts placement error from 2.88 to 0.20 positions for 10.9 questions instead of 5.6.
  • A simulated A/B test needs 481 users per arm to detect a 0.25-position improvement at 80% power, validated by 2,000 simulated experiments (A/A false-positive rate 4.35%).
  • 1,080 simulated transcripts replayed through the REST API give identical results; 8 PostgreSQL metric queries match pandas to 6e-14.

Code · Results

Fitted residual dynamics: transformers as interacting particles (research, in progress)

BruinML, with two collaborators; manuscript in preparation. Records the residual stream of pretrained models (GPT-2, BERT, ALBERT), fits an autonomous attention ODE over windows of layers, and tests the clustering, energy and metastability predictions of the Geshkovski–Rigollet theory on real checkpoints.

Real-time EEG focus detection (2025–2026)

CruX Neuroscience. Streams live 8-channel EEG at 250 Hz through notch and band-pass filtering and artifact rejection into a per-user calibrated workload index (frontal θ / parietal α), exposed through a Python API.

Multimodal signals for event-contract trading (2026)

Founder, UCLA Startup Labs. A scraping and embedding pipeline (custom CNN for images, adapted NLP model for text) maps social signals into a 4-D feature space; backtested on about 50 Kalshi markets over four months (Sharpe 1.38).

Experience

  • Researcher & Lead Developer, UCLA Math + Code + Art (Apr 2026–present)
  • Developer, VoroNav autonomy stack, independent project (2026–present)
  • Researcher, BruinML (Sep 2026–present)
  • Developer, PairRank, independent project (2026)
  • Developer, CruX Neuroscience (Sep 2025–Jun 2026)
  • Founder, UCLA Startup Labs (2026)
  • Founder, Twin Digital Media (2023–2025): built and sold a social-media marketing company at 18

Off the clock

Field and jazz band: SCSBOA field-band medals, a 4A Division first place, and the Louis Armstrong Jazz Award (2025). An IB research paper used Fourier analysis to compare the harmonic makeup of instrument tones.

Contact

Email winxonnguyen44@gmail.com · Résumé (PDF) · GitHub · LinkedIn. Available Summer 2027 · Los Angeles, CA · open to relocation.