# Yang Zhao's Resume

> Master's graduate in Software Engineering, with 3 years of foundational software development experience. Strong command of C/C++/Rust across embedded and Linux microservice work, with hands-on research in large language models and deep learning. Based in Qingdao; focused on backend engineering.

This file is an LLM-readable summary of this site. The full interactive resume (English): <https://cv.paulden.site/en/>; Chinese version: <https://cv.paulden.site/zh/>

## Basics

- Name: Yang Zhao
- Location: Qingdao
- Focus: Backend Engineer
- Resume (English): <https://cv.paulden.site/en/>
- Resume (Chinese): <https://cv.paulden.site/zh/>

## Summary

- Primary working languages are C/C++ and Rust, with solid Python and TypeScript.
- Practical experience with multithreading, asynchronous programming, RTOS development, fork/poll, and common network-programming models.
- A systematic understanding of large language models, covering concepts such as context window, token, and embedding, and familiarity with pre-training and post-training methods.
- Comfortable with containerization technologies such as Docker and Kubernetes: Docker for containerized development environments at work, and Kubernetes for containerized CI deployment.
- Familiar with agile development, CI/CD, and GitHub collaboration; good at boosting productivity with new tools, with a track record of pull requests and issues on open-source projects.
- Reads English documentation without difficulty, with working spoken and written English (CET-6 561).

## Education

- [Nanjing University of Science and Technology](https://cv.paulden.site/en/) — M.Eng. in Software Engineering (full-time), 2020.09 - 2023.04. Research focused on deep-learning video understanding. Awarded a first-class graduate scholarship and published one conference paper and one journal paper.
- [Hunan University of Science and Technology](https://cv.paulden.site/en/) — B.S. in Information and Computing Science (regular admission), 2016.09 - 2020.05. Served as class monitor and was named an outstanding student cadre; won second prize in the Hunan division of the Mathematical Contest in Modeling.

## Publications

- [Turning to a Teacher for Timestamp Supervised Action Segmentation](https://ieeexplore.ieee.org/abstract/document/9859626): IEEE ICME 2023, first author, Oral. Introduces a teacher model that acts as an "ensemble version" of the segmentation model during training, stabilizing pseudo labels and suppressing noise. Also designs a "segment-smoothness loss" that constrains more specifically how predicted probabilities transition smoothly within a single action segment.
- [Dilation-Erosion for Single-Frame Supervised Temporal Action Localization](https://link.springer.com/article/10.1007/s11042-023-15196-1): Multimedia Tools and Applications, fourth author. Designs a snippet-classification model and a dilation-erosion module: a dilation strategy first loosely expands candidate action segments to ease incomplete actions, then an erosion-like step removes background from those segments to reduce false detections.

## Experience

- **Qingdao Haier Washing Machine Co., Ltd. · Embedded Development Engineer** (2025.08 - Present; C/C++, Embedded, Intelligent voice assistant)
  Integrating LLM with a voice assistant; responsible for the embedded development and maintenance of the intelligent voice assistant.
  - Contributed to RTOS-based camera management and audio task scheduling: camera power-up/power-down management and photo capture management, plus audio focus management.
  - Contributed to voice-assistant troubleshooting: reducing wake-up latency and investigating memory leaks.
  - Contributed to the design of the team's internal CI setup: managing the CI task lifecycle with Kubernetes.
  - Participated in researching the integration of LLM with the voice assistant.
- **Nanjing ZTE Software Co., Ltd. · C/C++ Development Engineer, Core Network Platform** (2023.06 - 2025.08; C/C++, Rust, Linux, Microservices)
  Linux systems programming for the core-network platform, covering an intelligent operations service, an FTP-like file transfer protocol, and a stretch of CSV processing done in Java. Toolchain: GCC 4.8.5 and 12.2; version control with Git/Gerrit.
  - Intelligent O&M (C/C++, CMake, Docker, Kubernetes): collects network status across 5G services and detects anomalies, helping operations find problems in time and avoid serious network incidents.
  - FTP-like file transfer (Rust, shipped as a .so): rewrote inter-node file transfer (TIPC protocol) from C to Rust. Implemented chunked sending; for failed sends, a sliding window supports timeout retransmission and out-of-order delivery, and handling of mixed C and Rust programming.
  - CSV processing (Java): as an extension of intelligent O&M, joins the produced CSVs. Had no prior Java experience, got up to speed quickly and improved throughput with multithreading.
  - Responsible for cross-compilation and CMake builds on arm64, x86, and riscv64; strong command of CMake and the compile-and-link process.

## Skills

Microservices, Multithreading, Network programming, Embedded, Asynchronous programming, C/C++, Rust, TypeScript, Python, SQLite, CMake, Make, GDB, Docker, Kubernetes, Git, Ninja/n2, Cargo, Linux

## Articles

- [How Large Models Are Trained](https://mp.weixin.qq.com/s/qEdxbZNo1ZjLTGj10e7fmA): The training pipeline of large models: what pre-training learns, and what post-training (SFT and RLHF) sets out to solve.
- [What Is a Token in a Large Model](https://mp.weixin.qq.com/s/mLnTYgDdHhVu8vP5YKZFjQ): Explains what a token is and how tokenizer algorithms work.
- [How a Large Model Predicts Its Answers](https://mp.weixin.qq.com/s/14YorkYBjHU9Vf7HbwgbCA): How a large model predicts the next token from the context one at a time, and how sampling strategies and temperature influence the final output.

## Open-source Projects

- [oven](https://oven.paulden.site/): A coding harness written in Rust. Supports Windows, macOS, and Linux, and provides capabilities such as MCP, skills, memory, and subagents.
- [cargo-q](https://crates.io/crates/cargo-q): A small Cargo tool that runs several Cargo commands at once: cargo q fmt test build, to speed up development.
- [tab archive](https://marketplace.visualstudio.com/items?itemName=guza.tabarchive): A VSCode extension that closes long-open tabs on a schedule to reduce mental overhead.
- [numbers](https://github.com/guuzaa/numbers): A C++ component for handling integer addition and subtraction overflow. Implements Rust's four semantics — checked, overflowing, saturating, and wrapping — and works with GCC, Clang, and MSVC.