# Yufei Ma

AI Engineer at the U.S. Department of the Treasury.

# Identity
- Role: AI Engineer · U.S. Treasury
- Focus: Applied AI / ML systems
- Education: B.S. Computer Engineering
- Based in: San Diego, CA

# Projects
## tiktok-dlp
Self-hosted TikTok downloader and private archive with Discord monitoring, resumable creator imports, and a mobile-first video feed.
Stack: Node.js / Next.js / SQLite / yt-dlp / Docker / Cloudflare
- Built a bounded, deduplicated download pipeline for Discord commands, DMs, creator monitoring, photo posts, and tokenized large-file delivery.
- Designed full-profile imports with per-video checkpoints, restart recovery, cancellation, retry, duration limits, and restorable trash backed by SQLite and filesystem storage.
Source: https://github.com/nqrwhal/tiktok-dlp

## Postmaster
End-to-end tracking platform that unifies USPS, UPS, and FedEx behind one API and a power-user web dashboard.
Stack: TypeScript / Cloudflare Workers / Pages / KV / OAuth / Turnstile
- Designed and shipped a typed REST API on Cloudflare Workers that normalizes three carriers' tracking responses behind a single endpoint with consistent status, ETA, and event schema.
- Engineered the full request pipeline: OAuth client-credential auth per carrier, key-value caching with status-aware TTLs, per-IP rate limiting, and structured error handling.
Source: https://github.com/nqrwhal/postmaster

## Transformer-Based Music Generation
From-scratch PyTorch transformer that produces coherent multi-track symbolic music without melody prompts.
Stack: PyTorch / CUDA / POP909 / Hugging Face / BPE
- Trained an unconditional symbolic music transformer on the POP909 dataset, applying segmented tokenization to capture phrase-level musical structure across multiple instrument tracks.
- Built a CUDA-accelerated inference loop with configurable temperature and top-k sampling, surfacing the trade-off between conservative continuations and more exploratory outputs.
Source: https://github.com/nqrwhal/Transformer-Based-Music-Generation-Model

## Kendo Training Analysis Server
Dual-sensor IMU pipeline that grades kendo technique on form, force, and exertion in real time.
Stack: Python / Flask / SQLite / NumPy / WearOS / Raspberry Pi
- Led the backend: a Flask service on Raspberry Pi that ingests time-series IMU and heart-rate data from a Galaxy Watch and a paired shinai sensor, persists sessions to SQLite, and serves a dashboard for review.
- Designed strike-detection and form-scoring algorithms in NumPy — peak detection on accelerometer magnitude with 200ms refractory windowing, plus straightness and consistency scores derived from trajectory variance and waveform correlation.
Source: https://github.com/hongyuejin/UCSD-CSE-118-Team-1

# Current role
## U.S. Department of the Treasury — AI Engineer
Present · United States
AI Engineer at the U.S. Department of the Treasury.


# Technical capabilities
- Languages: Python, TypeScript, JavaScript, SQL, C++
- Application: React, Next.js, Node.js, FastAPI, Flask, Playwright
- Data + ML: PyTorch, CUDA, Hugging Face, Pandas, NumPy
- Systems: Cloudflare Workers, AWS, GCP, Docker, GitHub Actions, PostgreSQL, SQLite, Redis

# Contact
- Email: mailto:mayufei2004@gmail.com
- LinkedIn: https://linkedin.com/in/yufei-ma
- GitHub: https://github.com/nqrwhal