About
I work on machine learning and systems, and I like turning research prototypes into usable products.
I studied computer science at USC and start the UCLA M.S. in Computer Science in Fall 2026. In summer 2026, I worked on recommendation systems and edge AI at PMAOS while completing two paper submissions.
Applied ML internship
AI Engineer Intern. My main work was algorithm design for recommendation and multimodal inference under tight device constraints: structured interest propagation, multi-horizon diversity control, evidence-aware fusion, and fixed-point decisions that can abstain. A smaller research track covered millirobot screening and experiment design.
I designed recommendation logic that propagates preference signals in both directions between users and videos. The system keeps learned preference, retrieval scores, and diversity annotations separate, then allocates exposure across session, bridge, and long-term interest budgets. Fatigue controls reduce repetition without removing exploration.
I designed an audio-visual affect pipeline for 11 languages. It rejects misaligned audio–video pairs, applies speech PAD, and fuses the remaining evidence before a quantized Q15 decision layer. The final decision can abstain or expose conflicting modalities instead of forcing a label.
I compared untethered piezoelectric robot designs under 10 mm and selected a 9 mm four-channel candidate for further testing. I then designed a PZT–steel cantilever bench experiment with pass/fail criteria. The internship ended before procurement or prototype fabrication.
Papers and lab work
The paper adds a supervised router after the vision backbone to classify skin lesions with fewer tokens. Training constrains lesion retention, token budget, and routing diversity; the evaluation checks which lesion evidence remains on ISIC 2019.
The method encodes mass-conserving fragment pairs as sparse tokens before spectrum retrieval. Hard negatives and witness-removal tests check whether the exported evidence changes the ranking.
Selected informative key intervals before attention, implemented the method in Triton, and benchmarked it against HyperAttention and LevAttention.
arXiv:2505.11040 ↗Built reproducible RDKit, Chemprop, and hybrid-model benchmarks across seven Therapeutics Data Commons datasets, followed by feature-importance and SHAP analysis.
Personal projects
I built Fridge Manager with Next.js and Supabase. It uses current inventory, leftovers, recipes, dietary preferences, and available time to plan meals and grocery purchases.
howtoeat.food ↗This project connects a WeChat Mini Program, a browser workspace, knowledge retrieval, and FastAPI services for clinical and teaching workflows.
I built this iOS app to turn business-card photos into searchable bilingual records. It uses vision parsing first and falls back to OCR when needed.
Source code ↗Previous roles
Led an AI-assisted housing debt-collection product, including data pipelines, workflow automation, product roadmap, and engineering delivery.
Built Shopify-connected pages and Google Maps features; improved responsive behavior, ADA accessibility, SEO, and maintainability.
Built an LLM-assisted analysis pipeline for 5,000+ medical-malpractice documents and wrote laboratory workflow checks.
Gathered requirements for an internal medical-device platform and helped structure the program and hiring needs.
Support MedTech deal screening, market research, due diligence, financial modeling, and valuation.
Validated firmware and prototyped an ESP32-to-cloud I²C/MQTT telemetry path.
Education and tools
UCLA
M.S. Computer Science · starting Fall 2026
USC
B.S. Computer Science · 2026
GPA 3.76/4.00 · Viterbi Dean’s List
ML & systems
Python, C/C++, CUDA, Triton, PyTorch, TensorFlow
Product engineering
FastAPI, PostgreSQL, TypeScript, React, Next.js, Docker, Linux, AWS, GCP, MQTT