对技术有极致追求，对产品有较高审美 —— 追求极致的好看与好用。
Relentless about engineering excellence, obsessed with product aesthetics — built to be beautiful and usable.
中文
我是一名横跨 AI 算法、AI Infra 与机器人系统 的工程师，习惯从「算法 → 框架 → 算子 → 部署 → 产品」端到端地把事情做成。
- 🔬 AI 算法：自动驾驶感知、视频理解、行为检测
- ⚙️ AI Infra：深度学习训练框架、算法框架、大模型训练框架、推理引擎、算子开发
- 🤖 应用落地：全栈机器人研发（数据 → 模型 → 部署 → 机器人控制）
- 🎨 工程审美：既是技术极客，也是产品狂，追求极致的好看与好用
我相信优秀的系统 = 扎实的底层 + 优雅的抽象 + 克制的产品设计。
English
I'm an engineer working across AI algorithms, AI Infra, and robotics systems, comfortable owning the full path from algorithm → framework → kernel → deployment → product.
- 🔬 AI Algorithms: autonomous driving perception, video understanding, action detection
- ⚙️ AI Infra: deep learning training frameworks, algorithm frameworks, LLM training frameworks, inference engines, kernel development
- 🤖 Applied Systems: full-stack robotics (data → model → deployment → robot control)
- 🎨 Engineering Taste: a technical perfectionist and product fanatic — chasing the极致 of beauty and usability
Great systems = solid foundations + elegant abstractions + restrained product design.
AI 算法 / AI Algorithms
AI Infra
机器人 / 应用 / Robotics & Applications
工程 / 工具 / Engineering & Tools
|
|
|
|
-
OneNeuralNetwork — 中文：跨芯片平台的算子集合与统一神经网络库，后端基于 TVM / Triton 等深度学习编译器，目标是在异构芯片上运行 OpenMMLab、Detectron、DeepSpeed、Megatron-LM 等上层应用。
EN: A cross-chip operator collection and unified neural network library, built on DL compiler backends (TVM, Triton) to run OpenMMLab / Detectron / DeepSpeed / Megatron-LM on heterogeneous platforms. -
OneTensor — 中文：极简 Tensor 库，专为手写 CUDA kernel 时快速构造测试数据而设计，支持 Tensor 创建、Host/Device 数据搬运、npy/npz 加载与数据显示，相比 PyTorch/Cute 更轻量。
EN: A minimal Tensor library for quickly creating test data when writing CUDA kernels — Tensor creation, host/device transfer, npy/npz loading, and data view, lighter than PyTorch/Cute. -
RLink — 中文：面向分布式强化学习系统的轻量高性能通信层，解耦 Actor 采样与 Learner 训练，优化轨迹与模型参数的传输延迟。
EN: A lightweight, high-performance communication layer for distributed RL, decoupling Actor sampling from Learner training and optimizing trajectory/model-parameter transfer latency. -
pig-solver — 中文：一个 tiny 版本的 PyTorch。
EN: A tiny PyTorch — a minimal deep learning framework built for learning and hacking. -
UniRobot — 中文：一个具备训练和真机部署的全栈 robot 工程框架。
EN: A full-stack robotics engineering framework covering both training and real-world deployment.
“把复杂留给自己，把好用留给用户。”
“Keep the complexity for myself, leave the usability to the user.”