Biography

Research, built
into practice.

Personal AI, graph data systems, and confidential computing. From research to the systems people use.

English

About me

I think the most useful AI will be personal: it knows how you work, runs on hardware you own, and answers to you. That’s what I’m building as CTO of Siltok, with personal AI devices and the on-device inference system underneath them.

I got here from the bottom of the stack. I spent my Ph.D. at Tsinghua making Intel SGX practical (Occlum), led the development of AtlasGraph, a distributed graph database that ranked #1 in the world on the LDBC benchmark, and in 2025 founded AriseOS, a personal agent that learned your habits and workflows.

Research

Since moving to industry my research has followed the products I build. At Haizhi I first led AtlasGraph, a cloud-native distributed graph database, to #1 on the LDBC benchmark. Building on that, I took graph + LLM fusion from concept to product, using knowledge graphs to make LLM systems more factual and better at reasoning, and wrote it up in Big Data Research (2025). I helped define the LDBC Financial Benchmark (VLDB 2025), an industry-standard workload for graph databases in finance, and kept publishing with academic collaborators on data systems: vector query engines (SIGMOD 2025) and graph compression for resource-constrained analytics (VLDB 2024).

Before that, my Ph.D. work was about making Intel SGX practical for real programs. The main result is Occlum, a library OS that runs multiple processes inside a single SGX enclave, several to dozens of times faster than previous approaches (ASPLOS 2020), and the paper won Ant Group’s Best Paper of the Year. Occlum then grew well beyond the lab:

  • it became the core of SOFAEnclave, Ant Group’s trusted computing environment for financial workloads, and is named as one of the core technologies in Ant Group’s IPO prospectus;
  • it was donated to the Linux Foundation’s Confidential Computing Consortium (CCC), the first CCC open-source project initiated from China;
  • it is one of the recommended open-source runtimes for Microsoft Azure Confidential Computing.

Full list on DBLP.

Work

  • 2026–now Siltok (硅基词元), CTO: personal AI devices and on-device AI inference.
  • 2025 AriseOS, founder: built a personal agent that learned how you work by watching you do it once. Its memory was a knowledge graph on a graph database, mapping your tools, workflows and decision patterns, so it could act the way you would.
  • 2021–2025 Haizhi XingTu, Beijing: led the development of AtlasGraph, a cloud-native distributed graph database written in Rust, from architecture and team building to customer delivery in finance, telecom and energy. Under my lead, AtlasGraph ranked #1 in the world on the LDBC SNB Interactive benchmark in December 2023, beating the previous record by 45%. I also led graph + LLM fusion from concept to application system: proposing the idea, then building GraphAgent, which combines knowledge graphs with LLMs for enterprise AI.
  • 2019–2020 Ant Group, research intern: started and led the Occlum project.
  • 2018–2019 Intel Labs China, research intern: designed the multi-domain software fault isolation (MMDSFI) behind Occlum.
  • 2016–2017 VMware, intern: process-level virtualization for containers, KVM driver reuse.
  • 2014 Google Summer of Code: paravirtualized the PoK kernel to run RTEMS.

Education

  • 2015–2021 Tsinghua University, Ph.D. in Computer Science: systems security and trusted computing.
  • 2011–2015 University of Electronic Science and Technology of China (UESTC), B.Eng. in Embedded Systems.

Honors

  • 2025, Zu Chongzhi Award, AI Frontier Innovation Award (Outstanding Achievement of the Year), as a core contributor to a graph-and-LLM application platform built on large-scale graph analytics (Haizhi XingTu & Tsinghua University).
  • Ant Group Best Paper of the Year, for Occlum (ASPLOS 2020).

Mentoring, talks & press

  • Industry mentor at Renmin University of China (RUC).
  • 2024, People’s Daily: AtlasGraph’s results on the international graph database benchmark.
  • 2024, CCF Xiuhu Conference (large-scale graph computing and intelligent systems): talk on the feasibility and open challenges of fusing knowledge graphs with LLMs at the knowledge-base level.
  • 2023, Rust China Conf talk: AtlasGraph: a high-performance distributed graph database built in Rust.
  • 2023, People’s Daily: featured in a story on industry–academia collaboration and domestic graph databases.

Selected publications

  • 图模融合：人工智能系统事实表达和逻辑推理增强 (Graph–LLM fusion: enhancing factual expression and logical reasoning in AI systems). Juan Yang, Youren Shen. Big Data Research (大数据), 2025. paper
  • The LDBC Financial Benchmark: Transaction Workload. Shipeng Qi et al. VLDB 2025. paper
  • Tribase: A Vector Data Query Engine for Reliable and Lossless Pruning Compression using Triangle Inequalities. Qian Xu et al. SIGMOD 2025. paper
  • Improving Graph Compression for Efficient Resource-Constrained Graph Analytics. Qian Xu et al. VLDB 2024.
  • Occlum: Secure and Efficient Multitasking Inside a Single Enclave of Intel SGX. Youren Shen et al. ASPLOS 2020. paper
  • To Isolate, or to Share? That is a Question for Intel SGX. Youren Shen et al. APSys 2018. paper

Life

Off the keyboard I play tennis, ski in winter and dive when I can get to the sea.

中文

关于我

现任硅基词元 CTO，负责个人 AI 设备与端侧推理系统，目标是让 AI 运行在你自己拥有的设备上，并且只对你一个人负责。

我的技术积累集中在系统底层。博士期间在清华研发 Occlum，使 Intel SGX 能够运行真实程序；此后主导研发分布式图数据库 AtlasGraph，在 LDBC 基准测试中取得世界第一；2025 年创办 AriseOS，推出一款能学习并记住你工作方式的个人 Agent。

在我看来，最有用的 AI 是「个人的」：它了解你的工作方式，运行在你自己的设备上，也只对你一人负责。

研究

在海致星图期间，我先主导研发云原生分布式图数据库 AtlasGraph，在 LDBC 基准测试中取得世界第一；随后在图数据库的基础上做图模融合，借助知识图谱提升大模型系统在事实表达与逻辑推理上的可靠性，并推动其落地为产品。相关工作发表于《大数据》2025 年第 1 期。

我也参与制定了面向金融行业的图数据库基准 LDBC Financial Benchmark（VLDB 2025），并与高校合作开展数据系统方向的研究：向量查询引擎（SIGMOD 2025）、面向资源受限场景的图压缩（VLDB 2024）。

博士期间，我主要解决一个问题：让 Intel SGX 能够运行真实程序。成果是 Occlum，一个能在单个 SGX enclave 内运行多进程的库操作系统，性能较此前方案提升数倍至数十倍（ASPLOS 2020），论文获评蚂蚁集团年度最佳论文。此后，Occlum 走出实验室：

  • 成为蚂蚁可信计算环境 SOFAEnclave 的核心，并作为核心技术之一写入蚂蚁集团科创板招股书；
  • 捐赠给 Linux 基金会旗下的机密计算联盟（CCC），成为 CCC 首个由中国发起的开源项目；
  • 成为微软 Azure 机密计算推荐使用的开源运行环境之一。

工作

  • 2026–今 硅基词元，CTO：负责个人 AI 设备与端侧 AI 推理系统。
  • 2025 AriseOS，创始人：研发个人 Agent，能学习并记住用户的工作方式。其记忆层是构建于图数据库之上的知识图谱，记录用户使用的工具、工作流程与决策习惯，从而按用户的方式行事。
  • 2021–2025 海致星图：主导研发云原生分布式图数据库 AtlasGraph（Rust 实现、存算分离），涵盖架构设计与团队组建，直至金融、运营商、电力等行业的客户落地。2023 年 12 月，该产品在我主导下取得 LDBC SNB Interactive 性能世界第一，较原纪录提升 45%。此外，我主导了图模融合从概念到应用系统的研发，推出结合知识图谱与大模型的 GraphAgent。
  • 2019–2020 蚂蚁集团，研究实习生：发起并主导 Occlum 项目。
  • 2018–2019 英特尔中国研究院，研究实习生：设计 Occlum 底层的多域软件故障隔离（MMDSFI）。
  • 2016–2017 VMware，实习生：从事面向容器的进程级虚拟化与 KVM 驱动复用。
  • 2014 Google Summer of Code：将 PoK 内核半虚拟化，使其支持运行 RTEMS。

荣誉

  • 2025 年，第二届「祖冲之奖」人工智能前沿创新奖·年度突出成果奖，作为「基于大规模图数据分析技术的图模联合应用平台」项目核心参与人。
  • 蚂蚁集团年度最佳论文（Occlum，ASPLOS 2020）。

导师、演讲与报道

  • 担任中国人民大学企业导师。
  • 2024 年，人民日报报道了 AtlasGraph 在国际图数据库基准测试中的成绩。
  • 2024 年，在第十四期 CCF 秀湖会议（大规模图计算与智能系统）上作报告，从知识库层面探讨知识图谱与大模型融合的可行性及尚待解决的问题。
  • 2023 年，于 Rust China Conf 作演讲，题为《AtlasGraph：用 Rust 构建的高性能分布式图数据库》。
  • 2023 年，人民日报在一篇关于产学研融合与国产图数据库落地的报道中介绍了我的工作。

教育

  • 2015–2021 清华大学，计算机科学与技术博士，研究系统安全与可信计算。
  • 2011–2015 电子科技大学，嵌入式系统本科。

生活

工作之外，我打网球、冬季滑雪，有机会就去潜水。