Stefanie Nguyen personal website

Hello, I'm Stefanie — Hà Phương

If you want to know more about my journey, please visit this autobiography.


Me in a nutshell

I'm studying Math and Computer Science (Class of 2029) at Pomona College in Claremont, CA.

Right now, I'm interested in the intersection of pure math and practical ML/AI, with a soft spot for computer vision. Outside of research and coursework, I also just love building quirky, spontaneous projects simply for the fun of making things work.

If you want to yap or meet up or want to talk about anything, write to me !

Stefanie

My experience

7 stops so far, newest first — the short versions. The full stories live on my resume.

drag the timeline →

  1. Sep 2026 — Present

    Software Developer

    Associated Students of Pomona College (ASPC)

    Associated Students of Pomona College (ASPC) logo

    Software developer for Pomona's student government. Just getting started — more great things yet to come :)

  2. May 2026 — Present

    AI Program Fellow

    BREAK THROUGH TECH

    BREAK THROUGH TECH logo

    Twelve-month AI fellowship, selected from 4,300+ applicants. Building RAG pipelines and end-to-end predictive models, and iterating on them in reviews with industry AI professionals.

    PYTHONLANGCHAINSCIKIT-LEARNPYTORCH
  3. Mar 2026 — Present

    Student Researcher

    VIETNAM POLYMATH REU

    VIETNAM POLYMATH REU logo

    Research in Fourier analysis and signal recovery: proofs, exhaustive simulations, and a sole-author paper in preparation.

    FOURIER ANALYSISPYTHONPROOFS
  4. Mar 2026 — Present

    Academic Mentor

    PiMA

    PiMA logo

    Designed the linear algebra and convex optimization curriculum for the Math for Machine Learning track, and mentor a 4-student research team from first principles through their final presentation.

    CONVEX OPTIMIZATIONSVMMENTORSHIP
  5. Jan 2026 — Present

    Teaching Assistant

    POMONA COLLEGE CS

    POMONA COLLEGE CS logo

    Ran small-group mentor sessions and weekly labs for 30–40 intro CS students, with targeted feedback on code and problem-solving.

    TEACHINGPYTHONCODE REVIEW
  6. Oct 2025 — Present

    Research Assistant to Prof. Clark

    ARCS LAB @ POMONA

    ARCS LAB @ POMONA logo

    Sim-to-real robot navigation: training PyTorch vision policies on hybrid real/simulated data, building Unreal Engine tooling, and deploying to Jetson hardware with motion-capture evaluation.

    PYTORCHUNREAL ENGINEJETSON NANO
  7. Jul 2025

    Participant, Women in STEM

    JANE STREET

    JANE STREET logo

    Fully-funded Women in STEM program in Hong Kong on math and CS in quantitative trading — modeling outcome distributions and EV-maximizing strategies with traders and 50+ peers.

    PROBABILITYQUANT TRADINGGAME THEORY

Things I've built

Phở-Around

in progress · optimization, mip, llm, fastapi, react

A web-based itinerary planner that frames city wandering as an orienteering optimization problem. Powered by a PuLP/CBC mixed-integer linear solver, bilingual (EN/VI) schema-constrained LLM parsing, and a React 19 + FastAPI backend with offline travel-time routing.

P-resents

published · algorithms, next.js, tailwind

A gift-exchange planner built on bipartite matching: generates optimized rule sets from group traits, cutting manual matching time by 90%. Next.js + Tailwind, with RESTful APIs and secure authentication flows.


What I'm thinking about

[Working title] On exact recovery of binary signals from incomplete Fourier data

Phuong Nguyen and Thang Pham · Manuscript in preparation · Vietnam Polymath REU · 2026

If some of a signal's Fourier coefficients go missing, when can the signal still be reconstructed exactly? The guarantees in the literature are sufficient conditions — they certify recovery but leave a gap to where recovery actually stops. This paper lives in that gap: we believe the right answer is sharper than what counting arguments can see, and cheap to check. Details when the preprint is out.

[arXiv][pdf]publication coming soon!

How Much Real Data? Training-Data Composition for Sim-to-Real Visual Navigation on a Low-Cost Robot

with the ARCS Lab, Pomona College · Manuscript in preparation · 2026

How much real data does a deployable vision policy actually need? On a JetBot line-follower scored millimetre-level against motion-capture ground truth, we sweep real/simulated training mixes with the simulated set held fixed: neither source transfers alone, hybrids keeping at least 90% of the real pool complete every run at ~16 mm, and below that lies a cliff. The real fraction — not the absolute count of real frames — is what carries transfer.

[arXiv][pdf]publication coming soon!


Off the clock

Film photos from home, and a little web darkroom called Ham Chơi. Everything on the miscellaneous shelf happens off the clock, and still shapes how I build things.