Sandbox
隔离 Agent 工具执行，管理资源边界与沙箱生命周期。Isolate tool execution and manage resources and sandbox lifecycles.
INTERVIEW INTRODUCTION · 05 MIN
冉熙Xi Ran 中国科学院大学 · 2027 届计算机技术硕士UCAS · M.S. Computer Technology, 2027
把模型能力做成能在真实业务里稳定运行的 Agent 系统。
I turn model capabilities into reliable Agent systems for real-world operations.
国科大杭高院 × 中科院计算所 × 浙江大学联合培养
Joint training across HIAS, ICT and Zhejiang University
Agent InfraRuntime · Sandbox · Session
自动决策Decision SystemsTool Use · Workflow · Data
工程交付DeliveryEvaluation · QA · Open Source
面试官您好，我叫冉熙，是中国科学院大学计算机技术专业 2027 届硕士。我的核心定位是 Agent 工程：我关心的不只是模型能不能回答，而是它能否接入工具和数据，在真实业务里稳定运行，并对最终结果负责。
Hello, I am Xi Ran, a 2027 master's candidate in Computer Technology at UCAS. My focus is Agent engineering: not only whether a model can answer, but whether it can use tools and data reliably in real operations and be accountable for outcomes.
01 / TRAJECTORY
一条持续向系统底层深入的路径A path deeper into the system stack
02
从业务应用，到数据决策，再到 Agent 运行时，我一直在解决同一件事：
把复杂问题变成可运行的系统。
From applications to data-driven decisions and Agent runtimes, I solve one recurring problem: turning complexity into working systems.
主线：问题抽象 → 工具编排 → 系统验证Pattern: frame → orchestrate → verify
2023.10
趣链科技 · 区块链研发Hyperchain · Blockchain R&D从业务需求到全栈交付From requirements to full-stack delivery
趣链科技 · 区块链研发Hyperchain · Blockchain R&D从业务需求到全栈交付From requirements to full-stack delivery
2025.06
浙江大学 · 量化系统Zhejiang University · Quant Systems多源信号、策略与风控平台Signals, strategy and risk platforms
浙江大学 · 量化系统Zhejiang University · Quant Systems多源信号、策略与风控平台Signals, strategy and risk platforms
2025.12
Bybit · Agent 风控算法Bybit · Agent Risk Algorithms将专家排查固化为自动系统Operationalized expert investigation
Bybit · Agent 风控算法Bybit · Agent Risk Algorithms将专家排查固化为自动系统Operationalized expert investigation
2026.06
华为 · Agent InfraHuawei · Agent Infra运行时、沙箱与云原生开源Runtime, sandbox and cloud native OSS
华为 · Agent InfraHuawei · Agent Infra运行时、沙箱与云原生开源Runtime, sandbox and cloud native OSS
#1本科专业排名Undergraduate rank
3.75硕士 GPAMaster's GPA
50+开源仓库Open-source repositories
这几段经历看起来跨区块链、量化和云原生，但技术主线是一致的。早期我做完整应用，后来把数据和算法接入真实决策流程，现在进一步关注 Agent 的运行时、隔离和状态管理。我的能力是在业务、算法和工程三层之间往返。
The domains vary, but the engineering trajectory is consistent. I started with complete applications, then connected data and algorithms to operational decisions, and now work deeper on runtime, isolation and state management.
02 / CASE STUDY · BYBIT
把依赖经验的排查，变成 4 分钟全市场系统Turning expert investigation into a 4-minute market-wide system
03
问题：异常持仓分散在多个账户，人工串行取数、关联和验证，难以规模化。
Problem: abnormal positions were distributed across accounts, making manual retrieval and verification difficult to scale.
4 MIN完成全市场扫描for a market-wide scan
10+累计识别异常团伙groups identified
8D多维风险评分risk scoring
关键不是堆规则，而是把人的判断路径拆成 Agent 可以调用、验证和短路的工具链。
The key was not adding rules, but turning human reasoning into tools an Agent could invoke, verify and short-circuit.
01全景扫描Scan发现异常集中Detect concentration
02群体聚类Cluster形成候选关系Build candidates
03行为验证Verify交叉验证信号Cross-check signals
04风险评分Score统一判断口径Standardize decisions
05结论生成Report输出可复核证据Produce evidence
算法：Algorithm: 将主观经验固化为可解释评分explainable scoring from expert knowledge
工程：Engineering: 并行取数、条件路由、失败恢复parallel tools, routing and recovery
我在 Bybit 最有代表性的工作，是把异常持仓排查从专家经验改造成自动系统。先将流程拆成扫描、聚类、行为验证、评分和报告五个阶段，再把每一步做成可组合工具。算法侧用多维信号降低误报，工程侧加入并行、条件路由和短路优化，最终四分钟完成全市场扫描，累计识别十多个异常团伙。
At Bybit, I converted expert-driven position investigation into an automated system. I decomposed it into scanning, clustering, behavioral verification, scoring and reporting, then made each stage composable. The result was a market-wide scan in four minutes with more than ten abnormal groups identified.
03 / AGENT INFRA · HUAWEI
让 Agent 跑得安全、可恢复、可扩展Making Agents safe, recoverable and scalable
04
在华为参与云原生 Agent Infra 开源研发，关注的不是一次调用，而是长时间、多租户、真实负载下的运行可靠性。
At Huawei, I contribute to cloud-native Agent infrastructure, focusing on reliability under long-running, multi-tenant and real workloads.
Session
支撑状态持久化、上下文恢复与长任务连续运行。Support persistence, context recovery and long-running tasks.
Scheduling
基于 Kubernetes 参与资源感知调度与弹性执行。Contribute to resource-aware scheduling and elastic execution on Kubernetes.
HUAWEI · CNCF从企业场景回到开源上游：From enterprise needs to upstream: Karmada · AgentCube · Go · Kubernetes
在华为，我把关注点进一步下沉到 Agent Infra。企业 Agent 不能只在一次 Demo 中成功，它需要安全执行工具、保存和恢复状态，还要在 Kubernetes 上被调度。我的工作覆盖 Sandbox、Session 和调度三个方向，并通过 Karmada、AgentCube 等社区参与上游协作。
At Huawei, I moved deeper into Agent infrastructure. Enterprise Agents need isolated execution, recoverable state and schedulability on Kubernetes. My work spans Sandbox, Session and scheduling, with upstream collaboration in projects such as Karmada and AgentCube.
04 / BUILD IN PUBLIC
不只会调用框架，也能把 Agent 做成产品Beyond frameworks: building Agents into products
05
OfferPilot ↗TypeScript · Multi-Agent · ASR
手写 Agent Loop，编排诊断、简历与 JD 匹配子 Agent；支持 Tool Use、语音输入、上下文预算和流式输出。A hand-built Agent loop orchestrating diagnosis, resume and JD matching with tools, voice input and context budgets.
21 ★完整面试诊断闭环diagnostic loop
SlidePilot ↗TypeScript · Playwright · QA
从叙事规划到页面生成，再用 DOM 与像素检测自动发现溢出、空白和低对比度问题。Plans, renders and then detects overflow, whitespace and contrast defects through DOM and pixel QA.
7 STAGES生成到验收generation to QA
zero2Agent ↗Agent Engineering · Tutorial
把 Agent 基础、框架评测、工程实践和面试项目沉淀成结构化教程，持续接受社区反馈。A structured curriculum spanning foundations, framework evaluation, engineering practice and an interview project.
80 篇186 ★ · 8 modules
公开数据截至 2026.07Public metrics as of 2026.07
开源项目是我验证方法和产品能力的第二现场。OfferPilot 证明我能不依赖 LangChain 手写 Agent Loop，并处理多 Agent、工具调用和上下文预算；SlidePilot 把生成和自动验收放进同一条流水线；zero2Agent 则把实践沉淀为八个模块、八十篇内容。它们共同证明我能完成从实现到验证再到传播的闭环。
Open source is where I validate both methods and product thinking. OfferPilot demonstrates a framework-independent Agent loop, SlidePilot puts generation and QA in one pipeline, and zero2Agent turns practice into a structured 80-article curriculum.
05 / ROLE FIT
我能带来的，不只是一个 Agent DemoWhat I bring goes beyond an Agent demo
06
从问题定义到系统实现，从结果验证到开源沉淀。
From problem framing to implementation, validation and open-source learning.
期待参与 Agent Runtime、工具编排、评测体系或 AI 应用工程，把模型能力转化为稳定、可复用、可观测的系统能力。
I want to work on Agent runtimes, tool orchestration, evaluation or AI application engineering, turning models into reliable, reusable and observable systems.
01
懂业务与算法Business + algorithms能把模糊问题转成指标与流程Turn ambiguity into metrics and workflows
02
能做系统工程Systems engineering从 Agent Loop 到 Runtime 与云原生From Agent loops to runtime and cloud native
03
对结果负责Outcome ownership用数据、QA 与开源反馈持续验证Validate through data, QA and feedback
总结来说，我的差异化不是某一个框架，而是能完成问题定义、工具和系统实现、结果验证以及复用沉淀的完整闭环。我希望下一份工作继续做 Agent Runtime、工具编排、评测或 AI 应用工程，把模型能力做成真正能长期运行的系统。谢谢。
My differentiator is not one framework, but the complete loop from problem framing through implementation, validation and reusable learning. I hope to continue building Agent runtimes, orchestration, evaluation or AI applications that operate reliably over time.
APPENDIX / Q&A
能力地图：从底座到业务闭环Capability map: infrastructure to outcomes
A1
Agent Infra
Go · Kubernetes · Docker
Sandbox、Session、资源调度、状态恢复Sandbox, sessions, scheduling and recovery
Agent Engineering
Agent Loop · Tool Use · Multi-Agent
条件路由、并发编排、上下文预算、流式输出Routing, concurrency, context budgets and streaming
Data & Algorithms
Python · SQL · Presto/Hive
异常检测、风险评分、数据管道与自动报告Anomaly detection, scoring, pipelines and reporting
Product Delivery
TypeScript · FastAPI · Vue
需求抽象、全栈实现、自动 QA 与开源反馈Problem framing, full-stack delivery, QA and feedback
这页用于追问时快速定位：我的能力可以分成 Agent Infra、Agent 编排、数据算法和产品交付四层。每一层都有真实项目，不是只停留在概念或课程学习。
This appendix maps my experience across infrastructure, orchestration, data and product delivery. Each layer is backed by implemented work.
APPENDIX / BACKGROUND
教育背景与研究能力Education and research capability
A2
中国科学院大学University of Chinese Academy of Sciences
计算机技术硕士 · 2024.08–2027.07M.S. Computer Technology · 2024.08–2027.07
GPA 3.75 · 学业二等奖学金（前 20%）Academic Scholarship (Top 20%)
浙江工商大学Zhejiang Gongshang University
信息管理与信息系统学士 · 2020.09–2024.07B.E. Information Management · 2020.09–2024.07
GPA 3.77 · 专业第 1 · 一等奖学金Ranked 1st · First-Class Scholarship
研究生数学建模国二等奖Graduate Modeling National 2nd Prize2024 · 队伍 1/3team role 1/3
“挑战杯”全国二等奖Challenge Cup National 2nd Prize2023 · 队长team lead
论文与技术写作Research and technical writingSCI / IEEE 论文 · 开源教程SCI / IEEE papers · OSS tutorials
如果面试官关注教育和研究背景，可以补充本科专业第一、硕士 GPA 3.75，以及数学建模和挑战杯两项全国二等奖。这些经历训练了我将开放问题形式化、用数据验证并清晰表达结论的能力。
My academic background includes ranking first in my undergraduate major, a 3.75 master's GPA and two national second prizes, which trained me to formalize open problems, validate them with data and communicate clearly.