Sailglobal AI | AI Engineer

Sailglobal AI | AI Engineer

Location: Shanghai

Experience: 2–6 years

About Sailglobal AI

Hiring across borders, running payroll, and relocating employees remain slow, highly manual processes. Every country has its own employment laws, tax regulations, and immigration procedures—and those rules change frequently. Companies and their HR service providers rely on local experts to navigate these requirements, causing operating costs to rise with every new country and every new hire.

Sailglobal AI is an AI-native global people operations company.

We provide end-to-end services for customers hiring internationally, covering employee onboarding, payroll, visas and immigration, employee benefits, and offboarding. AI agents execute the work, while our specialists provide final review and approval for critical, irreversible actions.

Our experts teach the system a workflow once, and those capabilities can be reused across similar cases. As the business scales, the marginal cost of supporting each additional country or customer continues to decline.

We are already handling real cases for real customers and are expanding our service coverage to more than 50 countries.

Our vision is to make it as easy for any company to hire, pay, and relocate employees anywhere in the world as it is in its home country.

About the Role

You will build, optimize, and evaluate AI agents that perform real operational work.

Your work goes beyond calling a large language model. You will build the complete system around it: the agent harness, tool-use capabilities, safety controls, and evaluations that measure whether changes actually improve performance.

You will make AI agents reliable, robust, and controllable in complex, real-world cross-border HR operations.

Responsibilities

1. Build AI Agents That Perform Real Work

Develop agents that read documents, fill out web forms, and execute multi-step HR workflows, with human approval required before critical final actions.

2. Build the Agent Harness

Build the mechanisms surrounding the LLM, including tool calling, context management, guardrails, failure retries, and termination rules, to ensure agents operate reliably.

3. Develop Agent Evaluations (Evals)

Create evaluation datasets and metrics to determine whether changes to models, prompts, or system logic genuinely improve agent performance.

Integrate evaluations into every system change to detect performance degradation and functional regressions.

4. Create Reproducible Testing and Training Environments

Turn real business workflows into repeatable test environments for validating agent execution, with the potential to support agent training in the future.

5. Optimize Cost and Speed

Select appropriate models and runtime parameters for each task’s complexity and requirements, balancing execution quality, model cost, and latency.

Requirements

1. 2–4 years of software engineering experience , including at least one year building LLM applications used by real users.
2. Experience independently building an agent harness , beyond single-prompt calls. You have implemented agent systems with state management, tool-calling loops, retries, timeout controls, structured outputs, and guardrails.
3. Hands-on experience with at least one agent framework, such as LangGraph, OpenAI Agents SDK, Claude Agent SDK, or PydanticAI . You can clearly explain which problems these frameworks solve and when you would choose not to use one.
4. Experience building evaluation systems , including labeled datasets, metrics tied to real failure modes, and checks that identify performance regressions before release.
5. An understanding of the limitations of LLM-as-Judge , with the ability to validate model-based evaluations against human annotations rather than relying entirely on model-generated scores.
6. Strong Python programming skills , with the willingness to examine execution traces and patiently investigate the root causes of agent failures.
7. Sound judgment about AI capabilities and limitations, including which tasks models can execute autonomously and which decisions must remain under human control.

Nice to Have

1. Reinforcement Learning and Agent Training

Experience developing reinforcement learning (RL) environments, reward functions, or simulators to train or evaluate agents performing browser interactions, tool use, or coding tasks.

2. Browser Automation and Computer-Use Agents

Experience building browser agents or computer-use agents, with familiarity with technologies such as Playwright, the Chrome DevTools Protocol, or accessibility trees.

3. Document AI

Experience with OCR, document layout parsing, or multilingual information extraction.

4. Model Fine-Tuning and Distillation

Hands-on experience fine-tuning open-source models or performing knowledge distillation.

5. Durable Workflows and Human-in-the-Loop Systems

Experience with durable workflow engines such as Temporal, or with building human-in-the-loop systems.

6. Public Technical Work

Publicly accessible work that demonstrates your technical skills, engineering judgment, and practical experience, such as:

• Technical articles or reports on agent evaluation systems.
• Open-source agent tools or frameworks.
• Technical blog posts or project write-ups about agent systems you have built.

These examples help us better understand your capabilities, engineering approach, and hands-on experience.

Sailglobal AI | AI工程师

地点：上海

经验：2-6年

关于Sailglobal AI

跨境招聘、薪资管理和员工调动仍然是缓慢且高度依赖人工的流程。每个国家都有其自身的劳动法、税收法规和移民程序，而且这些规则经常变化。企业及其人力资源服务提供商依赖当地专家来应对这些要求，导致每增加一个国家或每增加一名员工，运营成本都会上升。

Sailglobal AI是一家以人工智能为核心的全球人力运营公司。

我们为国际招聘客户提供端到端服务，涵盖员工入职、薪资管理、签证和移民、员工福利以及离职流程。AI代理执行各项工作，而我们的专家则对关键且不可逆的操作进行最终审核和批准。

我们的专家只需对系统进行一次工作流程培训，这些功能即可在类似案例中重复使用。随着业务规模的扩大，支持每个新增国家或客户的边际成本将持续下降。

我们已在为真实客户处理实际案例，并将服务覆盖范围扩展至 50 多个国家/地区。

我们的愿景是让任何公司都能像在本国一样轻松地在全球任何地方招聘、支付和调动员工。

职位概述

您将构建、优化和评估执行实际操作工作的 AI 代理。

您的工作不仅仅是调用大型语言模型。您将围绕它构建完整的系统：代理框架、工具使用功能、安全控制以及用于衡量变更是否真正提升性能的评估机制。

您将使 AI 代理在复杂的、真实的跨境人力资源运营中可靠、稳健且可控。

职责

1. 构建执行实际工作的 AI 代理

开发能够读取文档、填写网页表单并执行多步骤人力资源工作流程的代理，关键的最终操作需要人工审批。

2. 构建代理框架

构建围绕 LLM 的机制，包括工具调用、上下文管理、安全防护、故障重试和终止规则，以确保代理可靠运行。

3. 开发代理评估 (Evals)

创建评估数据集和指标，以确定对模型、提示或系统逻辑的更改是否真正提高了代理的性能。

将评估集成到每个系统更改中，以检测性能下降和功能回归。

4. 创建可复现的测试和训练环境

将真实的业务工作流程转化为可重复的测试环境，用于验证代理的执行情况，并有可能在未来支持代理训练。

5. 优化成本和速度

根据每个任务的复杂性和需求选择合适的模型和运行时参数，平衡执行质量、模型成本和延迟。

要求

2-4 年软件工程经验，包括至少一年构建供真实用户使用的 LLM 应用程序的经验。

具备独立构建代理框架的经验，而不仅仅是单次提示调用。您已实现过具有状态管理、工具调用循环、重试、超时控制、结构化输出和防护机制的代理系统。

至少具备一种智能体框架的实际操作经验，例如 LangGraph、OpenAI Agents SDK、Claude Agent SDK 或 PydanticAI。您能够清晰地解释这些框架解决的问题以及何时会选择不使用它们。

具备构建评估系统的经验，包括标注数据集、与实际故障模式相关的指标以及发布前识别性能退化的检查。

了解 LLM 作为评判者的局限性，能够根据人工标注验证基于模型的评估结果，而不是完全依赖模型生成的分数。

精通 Python 编程，并愿意检查执行轨迹，耐心探究智能体故障的根本原因。

对 AI 的能力和局限性有良好的判断力，包括哪些任务模型可以自主执行，哪些决策必须由人工控制。

加分项：

1. 强化学习和智能体训练

具备开发强化学习 (RL) 环境、奖励函数或模拟器的经验，用于训练或评估执行浏览器交互、工具使用或编码任务的智能体。

2. 浏览器自动化和计算机使用代理

具备构建浏览器代理或计算机使用代理的经验，熟悉 Playwright、Chrome DevTools 协议或辅助功能树等技术。

3. 文档人工智能

具备 OCR、文档布局解析或多语言信息提取方面的经验。

4. 模型微调和知识提炼

具备微调开源模型或进行知识提炼的实践经验。

5. 持久工作流和人机交互系统

具备使用 Temporal 等持久工作流引擎或构建人机交互系统的经验。

6. 公开技术作品

能够公开展示您技术能力的公开作品

技术技能、工程判断力和实践经验，例如： 关于智能体（Agent）评估系统的技术文章或报告。 开源的智能体工具或框架。 关于您构建的智能体系统的技术博客文章或项目总结。 这些示例有助于我们更好地了解您的能力、工程方法和实战经验。