Available for opportunities

Jerry Huang

Software engineer specializing in data engineering, machine learning infrastructure, and full-stack development. UCLA Mathematics of Computation graduate based in Los Angeles.

Beihai Park, Beijing, China — photo by Jerry Huang

Academic Background

University of California, Los Angeles
B.S. Mathematics of Computation
Relevant coursework: Machine Learning, Operating Systems, Applied Numerical Methods, Optimization, Software Construction, Probability & Statistics, Mathematics of Finance
Los Angeles, CA
March 2024

Work History

Quantitative Software Engineer Intern
Mind Fund Equity
Beijing, China
Jan 2025 – Jun 2025
  • Built a high-throughput GPU analytics pipeline integrating kdb+, Apache Arrow, and libcuDF, achieving 4× speedup over CPU-based processing across 50M+ daily trade records.
  • Implemented zero-copy IPC (kdb+ → Arrow → cuDF) to leverage CUDA vectorized operations and group-by aggregations; profiled with NVIDIA Nsight Systems, boosting GPU utilization from 58% to 87%.
  • Designed a results feedback loop returning GPU outputs to kdb+ for downstream analysis and storage, cutting end-to-end pipeline latency by 40%.
Software Engineer
Cal Kitchen
Los Angeles, CA
Sep 2022 – Present
  • Developed a Python inventory management system with a web dashboard, automating stock tracking, low-inventory alerts, and purchase order generation across 500+ SKUs.
  • Deployed a self-hosted SMB file server on Proxmox with RBAC, rsync nightly backups, and ZFS snapshots; managed 3× Dell PowerEdge 740XD nodes with 48 TB RAIDZ2 storage at 99.95% uptime.
  • Built and scaled a Shopify storefront with custom Liquid theme development, third-party integrations (payments, logistics, analytics), and automated fulfillment workflows, driving 3× revenue growth.
  • Oversaw trademark registration, product label and packaging design, and Cloudflare security (WAF, SSL/TLS, DNS), coordinating vendors to bring 4 SKUs to market.

Things I've Built

SceneSeek
ML

Semantic video retrieval system that indexes keyframes via CLIP (ViT-B/32) embeddings, supporting image-to-image and text-to-image search. Features a two-pass scene detection pipeline with coarse boundary detection and local refinement, diverse keyframe selection via MMR and semantic centrality, and hybrid retrieval reranking with concurrent S3 uploads and batched Qdrant vector upserts.

Python PyTorch CLIP Qdrant OpenCV Streamlit boto3 NumPy
OpenTrans
PyPI Open Source

Privacy-first local-LLM batch translation CLI that mirrors directory hierarchies into a target language, supporting Gemma, Llama 3, and DeepSeek via Ollama. Uses AST-based syntax shielding with tree-sitter to protect code blocks, LaTeX, frontmatter, and HTML/JSX. SHA-based incremental caching cuts redundant LLM calls by 80%; parallelized with ThreadPoolExecutor.

Python Ollama tree-sitter Click ThreadPoolExecutor YAML
DualAttenNet
ML

Dual-branch attention CNN for medical image classification with a global attention module for long-range inter-region dependencies and a local attention module for fine-grained pathological feature localization. Achieved 91.4% top-1 accuracy on a 22-class dataset, outperforming ViT-B/16 by 3.2%. Reduced parameter count 60% vs. ViT; cut training time from 14 hrs to 4.5 hrs via FP16 CUDA training. Validated with Grad-CAM heatmaps.

Python PyTorch CUDA NumPy Matplotlib Grad-CAM
Blacklist Reviews
Web

Community-driven business review platform with fuzzy search, interactive Leaflet.js map with geographic bounds filtering, leaderboard rankings, and multilingual support (English/Chinese). Includes a full moderation toolkit: review approval, user banning, content flagging, and admin dashboard. Built with custom Google OAuth2 (no django-allauth) and HTMX for dynamic interactions without a JS framework.

Django 5.2 HTMX PostgreSQL Google OAuth2 Leaflet.js Docker Gunicorn
Daily Grove · 打卡林
Web Open Source

Cozy pixel-art daily journaling app inspired by Stardew Valley. Multi-user, self-hostable Django app with color-coded categories, calendar heatmap, streak counter, markdown support with live preview, Chart.js stats, and fuzzy client-side search. Dark forest-green pixel aesthetic with VT323 and Press Start 2P fonts. AGPLv3 licensed.

Python 3.14 Django 5+ HTMX Alpine.js Tailwind CSS Chart.js SQLite Docker
LA Surcharge Explorer
Data

Interactive Dash app mapping Los Angeles restaurant surcharge data with a Mapbox choropleth, rich pin popups, fuzzy search, and bilingual (English/Chinese) UI. Features surcharge/rating/policy filters, live stats, dark mode, cookie consent, and Google Analytics. Includes server-side pagination, rate limiting, and anti-scraping protections. Deployable on Vercel or a self-hosted Gunicorn stack.

Python Dash Plotly Pandas Mapbox Flask-Limiter Vercel

Technologies

Languages

Python C++ R SQL Shell / Bash LaTeX

Frameworks & Libraries

PyTorch NumPy pandas scikit-learn Apache Arrow libcuDF Django HTMX Streamlit Dash

Infrastructure & Tools

Docker Git Linux Proxmox PostgreSQL CUDA NVIDIA Nsight Cloudflare SSH Jupyter STM32