about me.

Hi! I'm Veronica, a CS student at Stanford. I'm drawn to problems at the intersection of perception and understanding — computer vision, world modeling, and how intelligent systems represent and reason about the world. I've done research in surgical video understanding (phase recognition with transformers), quantum computing (adiabatic state preparation scaling), and dynamic graph learning. Outside of research I build full-stack products and spend time on graphic design and studio art. Currently mainly working in C++ and learning 3D rendering plus AI/ML.

education.

• Stanford University

  B.S. Computer ScienceStanford, CA

  2025 – 2029

experience.

• Incoming Systems Software Engineer Intern Seattle, WA

  at, boeing.com

  June 2026 – August 2026

  • Joining the Otter team to develop low-level mission systems software and a virtual integration platform for the Boeing 737 Next Gen and P-8 Poseidon maritime patrol aircraft.
  • Building simulated and virtualized mission system software that models hardware behavior across acoustics, radar, and vision systems — used for pilot training and software validation.
  • Improving testing and scripting infrastructure for debugging mission systems software; working in C, C++, Python, and bash at the OS level rather than application or sensor layers.

  • C
  • C++
  • Python
  • Bash
  • Perl
  • Mission Systems
  • Virtual Integration

• Research Author College Park, MD

  at, arxiv.org

  w/ Thomas D. Cohen (UMD), Hyunwoo Oh (UMD)

  May 2024 – October 2024

  • Accepted to European Physical Journal A: Hadrons and Nuclei (EPJA-108258.R2). Reviewer considered this to be "a very significant result."
  • Co-authored with physicists at the University of Maryland, providing numerical evidence for a conjecture about the computational cost of adiabatic quantum state preparation — a key step in quantum simulation of field theories.
  • Demonstrated that the cost proxy Q_D scales as L log L (superlinear) in path length L, confirming the conjecture that adiabatic state preparation is generically more expensive than linear-scaling alternatives as system size grows.
  • Proved a no-go theorem showing why wall-clock time is not a valid cost proxy: rescaling the Hamiltonian energy scale changes time but leaves errors unchanged, so a dimensionless quantity Q_D is required.
  • Studied random 4×4 Hamiltonians with time-reversal symmetry and non-periodic dynamics across path lengths spanning four orders of magnitude; compared three variants (Q_D1, Q_D2, Q_D1/2) and found consistent superlinear growth.

  • Python
  • Mathematica
  • Quantum Simulation
  • Adiabatic Theorem
  • Numerical Methods

• Assistive Technology Designer (Contract) Remote

  at, chimes.org

  October 2023 – January 2024

  • Designed an RFID-based audio identification device for visually impaired custodial employees at Chimes, a nonprofit employing people with disabilities — helping workers safely distinguish cleaning chemicals without relying on color or label recognition.
  • Built a two-ESP32 pipeline: an RFID reader identifies tagged equipment and transmits the ID via ESP-NOW to a receiver ESP32, which matches it to an audio file on an SD card and plays it through a 3D-printed speaker enclosure.
  • Ran a blindfolded user study with 6 participants across 6 trials — participants identified 5 chemicals 30.1% faster with the device; iterated on the design based on results, adding tactile attachment points and upgrading to an external speaker module.
  • CAD-modeled a custom speaker enclosure in OnShape with honeycomb mesh acoustics, easy-repair access, and structural independence; housed the receiver ESP32, serial MP3 module, and speaker driver inside.
  • Device was designed to scale to hundreds of employees across Chimes locations.

  • ESP32
  • RFID
  • ESP-NOW
  • C++
  • Arduino
  • OnShape
  • CAD
  • 3D Printing
  • Raspberry Pi

projects.

• researchml

  Surgical Phase Recognition for Aneurysm Clipping

  Stanford University · CS231N

  w/ Emily Oberleitner, Nicole Wong, Dr. Jinendra Ekanayake

  • F1@10: 0.944
  • Best Val Acc: 95.7%
  • Edit Dist: 0.809

  • Collaborated with Stanford School of Medicine.
  • Sourced a proprietary dataset of 48 intraoperative microscope videos (40,725 labeled frames) of aneurysm clipping surgery, annotated using CVAT across 4 surgical phases: Brain Exposure, Parent Vessel Identification, Dome & Neck Identification, and Clipping. Used video-level train/val splits to prevent temporal data leakage.
  • Designed NeuroOperA, a causal transformer for phase recognition adapted from the laparoscopic OperA — and outperformed it, achieving a Viterbi segmental F1@10 of 0.944 vs. OperA's ~0.80. Also implemented MS-TCN (which OperA did not), achieving ~95% validation accuracy vs. OperA's 92%.
  • Showed that fine-tuning ResNet50 on surgical frames (vs. frozen ImageNet weights) was the single largest factor: frame accuracy jumped from 53% to 95.7% and F1@10 from 0.553 to 0.895.
  • Applied Viterbi decoding with a data-driven learned transition matrix, outperforming hand-crafted surgical priors across all thresholds.

  • Python
  • PyTorch
  • ResNet50
  • MS-TCN
  • Transformer
  • Viterbi Decoding
  • CVAT
  • t-SNE
  • Confusion Matrix
  • Ablation Study
  • GCP
• researchml

  EvolveGCN-T: Self-Attention for Dynamic Graph Weight Evolution

  Stanford University · CS229

  w/ Victoria Yang, Kaci Morris

  • +8.4pt micro-F1 on Bitcoin-OTC
  • GRU → Transformer weight evolution
  • Reproduced baselines to ±1%

  • Proposed EvolveGCN-T, replacing EvolveGCN's GRU-based weight evolution with a Transformer encoder that self-attends over the explicit history of GCN weight matrices — the first method to apply attention directly to evolving weight sequences rather than node embeddings.
  • Outperformed the matched recurrent baseline (EvolveGCN-O) on Bitcoin-OTC edge classification: micro-F1 0.783 vs. 0.699, a +8.4 point improvement.
  • Reproduced published EvolveGCN baselines to within ±1% (Elliptic illicit-F1: 0.578 vs. paper's 0.51; SBM MAP: 0.194 vs. 0.199) before introducing the proposed variant.
  • Identified optimization instability — not context length — as the primary bottleneck; self-attention showed no consistent benefit from longer history windows, collapsing at h=10.
  • Evaluated across 3 benchmark datasets: Elliptic (Bitcoin fraud detection), SBM (synthetic link prediction), and Bitcoin-OTC (signed trust network edge classification).

  • Python
  • PyTorch
  • Graph Neural Networks
  • Transformer
  • EvolveGCN
  • Weights & Biases
  • Scikit-learn
  • Docker
  • GCP
• full-stackml

  JobShield: Detecting Fraudulent Job Postings

  w/ Yohannes Aklilu, Anna Roth, Anayochukwu Edwin Uche, Victoria Yang

  • F1: 0.913
  • 4.5× cheaper than LLM-only
  • 95/100 fraud caught

  • Built a full-stack job posting platform with a three-layer fraud detection pipeline targeting real malware attack vectors (OtterCookie, FlexibleFerret) that have been active since 2024.
  • Hybrid LR→LLM pipeline achieved F1 of 0.913, catching 95/100 fraudulent postings while being 4.5× cheaper than using Gemini alone ($0.077/1k vs $0.35/1k).
  • Automatic feedback loop injects every moderator decision as a labeled few-shot example into subsequent LLM calls, enabling continuous improvement without retraining.
  • Logistic Regression + TF-IDF on posting text plus 8 metadata flags (missing salary, missing requirements, etc.) handles obvious cases in under 0.05ms.

  • Next.js
  • TypeScript
  • Supabase
  • Gemini 2.5 Flash
  • Scikit-learn
  • TF-IDF
  • Vercel
  • PostgreSQL
  • GCP

technical skills.

• Languages:

  Python, JavaScript, TypeScript, HTML/CSS, LaTeX, C++, C, Java, x86

• ML & AI:

  PyTorch, Scikit-learn, ResNet50, Transformers, RNNs, Graph Neural Networks, MS-TCN, TF-IDF, Viterbi Decoding, Weights & Biases

• Web & Full-Stack:

  React.js, Next.js, Node.js, Tailwind CSS, ShadCN UI, Express.js

• Data & Backend:

  PostgreSQL, MongoDB, Supabase, Excel

• Cloud & Tools:

  AWS, Vercel, GCP, Git, Docker, CVAT

• Quantum & Scientific Computing:

  Adiabatic state preparation, Quantum simulation, NumPy, SciPy

• Engineering:

  OnShape, Fusion 360, 3D Printing, Woodworking, Soldering, Arduino, Raspberry Pi, CNC

• Design:

  Figma, Procreate, Procreate Dreams

art.

Figure

Oil on canvas · 2024

Turbulence

Oil on canvas · 2024

Oil Painters of America — Show #84271

Long-tailed Duck

Acrylic on gessobord · 2023

FWS Junior Duck Stamp — MD State runner-up, national traveling display

Forrest

Oil on canvas · 2024

Blue Marble Review — Issue 36

Untitled

Charcoal · 2025

White-cheeked Pintails

Oil on gessobord · 2023

FWS Junior Duck Stamp — MD State runner-up, national traveling display

Playdate

Colored pencil · 2021

Celebrating Art Magazine — High Merit

Greater

Colored pencil · 2022

Congressional Art Contest — Best Drawing, Poolesville HS

Triptych

Ink and digital · 2024

Paths

Oil on canvas · 2023

Blue Marble Review — Cover art

Peonies

Oil on canvas · 2022

Suzhou Canal

Oil on canvas · 2025

Garden Path

Watercolor · 2024

For Grandpa

Procreate · 2024

Metamorphosis

Ink and gelly pen · 2022

Slopes

Colored pencil · 2026

Forrest W.I.P

Acrylic · 2026

Pebbles

Oil on gessobord · 2022

Persimmon

Procreate · 2024

Face

Acrylic · 2022

Field

Watercolor · 2024

Still Life

Watercolor · 2023

Flower

Watercolor · 2021

W.I.P

Oil pastel · 2026

Blue

Watercolor · 2021

Canal W.I.P

Oil on gessobord · 2026