AI infrastructure · Inference optimization · Reliable AI applications

English · 中文 · Intel portfolio · WealthPortfolio

I'm a Computer Science and Mathematics undergraduate at The University of Texas at Austin, expecting to graduate in December 2027. Previously, I was a Software Engineer Intern at Intel (May-August 2026) and Inspur Enterprise Cloud (June-August 2025).

I work on making AI systems faster, measurable, and reliable: heterogeneous inference, model quantization, LLM serving, and AI agents with explicit transaction and audit boundaries. I'm interested in AI infrastructure, inference engineering, and software engineering opportunities in China and the United States.

OpenVINO · NNCF · ONNX Runtime · CPU / GPU / NPU

• PaddleOCR-VL 1.5: built isolated benchmarks and output-validation pipelines for INT8 quantization; measured approximately 2.8x end-to-end speedup and 27% lower memory usage versus the uncompressed OpenVINO baseline in the reported experiments.
• PP-OCRv6: rebuilt preprocessing, inference, and decoding; routed dynamic-shape detection to GPU and fixed-shape recognition to NPU with CPU fallback. ONNX Runtime parity testing helped diagnose a batch-correctness bug; throughput recovered from approximately 0.5 to 5.9 FPS compared with the initial faulty port.
• FLUX.2-klein-4B: brought up and profiled inference on Panther Lake, with per-stage latency, load-time, peak-memory, and version-regression tooling. In the documented 1024x1024, four-step, no-CFG test, INT4 reduced transformer peak RAM from 13.57 GB to 3.88 GB.
• Also built a SAM3 + LaMa object-removal demo with PyTorch/OpenVINO backends and a Gradio interface.

The portfolio contains public FLUX scripts and benchmark reports. Some linked OCR and super-resolution source repositories are private. Results are specific to the documented hardware, configurations, and test samples.

Next.js · TypeScript · FastAPI · PostgreSQL · Celery · RabbitMQ · Redis

A full-stack asset-management application that combines portfolio accounting with reviewable AI workflows.

• Transaction event ledger, double-entry journal postings, idempotency protection, row-level locking, and rebuildable holding projections.
• An auditable AI agent with typed tools, human confirmation and stale-state checks before mutations, and atomic multi-tool database writes.
• Asynchronous agent, market-data, and document jobs with persistent state and WebSocket updates.
• PDF/image processing, OCR, and document retrieval with page-level citations and reviewable transaction drafts.

• OS-vLLM: built a vLLM Prefill-Decode serving prototype on openEuler with OpenAI-compatible routing, health monitoring, session affinity, retries, and worker recovery. Evaluated 1x1 through 2x2 worker configurations using fault injection and latency/throughput benchmarks.
• Pintos: implemented preemptive priority scheduling, nested priority donation, synchronization, lazy loading, swap, page-fault handling, and clock-based page replacement.
• Inspur Enterprise Cloud: built a hospital appointment prototype using React/TypeScript, Node.js, Express, and PostgreSQL, including JWT authentication and transactional booking-conflict detection.

• Languages: Python, C/C++, TypeScript/JavaScript, Java, SQL
• AI and inference: OpenVINO, NNCF, ONNX Runtime, PyTorch, vLLM
• Applications and systems: FastAPI, React/Next.js, PostgreSQL, Redis, RabbitMQ, Celery, Docker, Linux

我是德克萨斯大学奥斯汀分校（UT Austin）计算机科学与数学本科生，预计于 2027 年 12 月毕业。曾于 2026 年 5-8 月在 Intel 担任软件工程实习生，并于 2025 年 6-8 月在浪潮企业云实习。

我关注 AI 系统的性能、可测量性与可靠性，主要方向包括异构推理、模型量化、LLM 服务，以及具备明确事务与审计边界的 AI Agent。希望寻找中国或美国的 AI 基础设施、推理工程及软件工程岗位。

OpenVINO · NNCF · ONNX Runtime · CPU / GPU / NPU

• PaddleOCR-VL 1.5： 搭建独立进程基准测试与输出一致性验证流程，使用 INT8 量化；在报告所述实验中，相比未压缩 OpenVINO 基线，实现约 2.8 倍端到端加速和 27% 内存下降。
• PP-OCRv6： 重建预处理、推理和解码流程，将动态形状检测分配至 GPU、固定形状识别分配至 NPU，并以 CPU 回退。通过 ONNX Runtime 对齐测试定位批处理正确性问题，吞吐量从初始错误移植版本的约 0.5 FPS 恢复至 5.9 FPS。
• FLUX.2-klein-4B： 完成 Panther Lake 平台适配，搭建分阶段延迟、加载时间、峰值内存与版本回归测试。在文档所述 1024x1024、4 步、关闭 CFG 的实验中，INT4 将 Transformer 峰值 RAM 从 13.57 GB 降至 3.88 GB。
• 构建 SAM3 + LaMa 目标移除演示，支持 PyTorch/OpenVINO 后端与 Gradio 界面。

项目集包含公开的 FLUX 脚本与测试报告；部分 OCR 和超分辨率源码链接指向私有仓库。性能结果对应各报告中的硬件、配置与测试样本。

Next.js · TypeScript · FastAPI · PostgreSQL · Celery · RabbitMQ · Redis

将投资组合账务与可人工复核的 AI 工作流结合的全栈应用。

• 交易事件账本、复式记账分录、幂等保护、行级锁及可重建的持仓投影。
• 类型化工具驱动的可审计 AI Agent：写操作前进行人工确认与状态过期校验，多工具写入在数据库事务中原子执行。
• Agent、行情与文档任务异步执行，保留任务状态并通过 WebSocket 更新进度。
• PDF/图片处理、OCR 与文档检索，返回页级来源引用及可复核的交易草稿。

• OS-vLLM： 在 openEuler 上构建 vLLM Prefill-Decode 服务原型，实现 OpenAI-compatible 路由、健康监控、会话亲和、重试和 Worker 恢复；通过故障注入及延迟/吞吐量基准评估 1x1 至 2x2 Worker 配置。
• Pintos： 实现抢占式优先级调度、嵌套优先级捐赠、同步机制、懒加载、Swap、缺页处理和 Clock 页面置换。
• 浪潮企业云： 使用 React/TypeScript、Node.js、Express 与 PostgreSQL 开发医院预约原型，实现 JWT 鉴权及基于事务的预约冲突检测。

• 语言： Python、C/C++、TypeScript/JavaScript、Java、SQL
• AI 与推理： OpenVINO、NNCF、ONNX Runtime、PyTorch、vLLM
• 应用与系统： FastAPI、React/Next.js、PostgreSQL、Redis、RabbitMQ、Celery、Docker、Linux