Building verifiable AI systems across model training, inference, agents, and data.
构建可验证的 AI 系统，关注模型训练与推理、Agent 和数据基础设施。
I work on open-source AI and systems projects, with an emphasis on reproducible experiments, explicit correctness checks, and usable engineering tools.
我主要做 AI 与系统方向的开源项目，重视可复现实验、明确的正确性验证和可实际使用的工程工具。
| Area / 方向 | Project / 项目 | What it does / 项目简介 |
|---|---|---|
| Inference / 推理 | RAFTInfer | RTX 50 系列上的 RAFT/RMM 推理运行时，结合 C++/CUDA 算子与 Rust CLI；以 Qwen3.5-9B BF16 的正确性和性能验证为边界。 RAFT/RMM-based inference for RTX 50 GPUs, with C++/CUDA operators and a Rust CLI; validation is scoped to Qwen3.5-9B BF16. |
| Generative recommendation / 生成式推荐 | QwenOneRec-0.8B | 基于 Qwen3.5-0.8B 的语义 ID 推荐实现，串起 SFT、约束生成、评估和 GRPO；公开验证覆盖最小端到端流程。 Semantic-ID recommendation with Qwen3.5-0.8B, connecting SFT, constrained generation, evaluation, and GRPO in a validated minimal flow. |
| Agent research / Agent 研究 | dsh-dream-rsi | DeepSeek Harness 的实验性插件，通过固定评测器、历史树回放和策略修订探索 coding agent 的闭环改进。 An experimental DeepSeek Harness plugin for coding-agent exploration using a fixed evaluator, historical-tree replay, and strategy revision. |
| Automated research / 自动化研究 | RAM-AutoResearch | 面向推荐、广告和营销的研究平台，组织实验、评估、诊断与报告。 A research platform connecting experiments, evaluation, diagnostics, and reporting for recommendation, advertising, and marketing. |
| GPU algorithms / GPU 算法 | RaftForge | C++20/CUDA 算法工具集，覆盖统计、模拟、图分析与优化；图和优化模块仍为预览阶段。 A C++20/CUDA toolkit for statistics, simulation, graph analytics, and optimization; graph and optimization modules are in preview. |
- Agents and prompts / Agent 与 Prompt： CareCopilot · AgentPlane · prompthelix · prompthelix-MultiLabel
- GPU and distributed systems / GPU 与分布式系统： cuvs-lab · DistAlgo · grpc-tfserving
- Data infrastructure / 数据基础设施： OpenEventFlow · Meta-AutoData
Top external repositories by merged PRs (up to 10). / 按已合并 PR 数展示外部仓库（最多 10 个）。
Counts are checked about every 15 minutes, subject to GitHub scheduling and search indexing. / 每约 15 分钟检查一次，实际更新受 GitHub 调度和搜索索引影响。All merged public PRs / 全部已合并的公开 PR.
GitHub's native contribution graph appears below this README on my profile. / GitHub 原生贡献图位于个人页的本段 README 下方。