强化学习算法工程师（Agent Reinforcement Learning Engineer）

**部门：**AI智能体研究中心
**地点：**香港, 深圳
**工作经验：**实习生、应届生
**招聘人数：**10人

关于岗位

我们正在招聘 AI 智能体强化学习工程师，加入 echOS Agent 核心团队。

你将直接参与构建在真实产业环境中运行的智能体系统，让 Agent：

• 与复杂业务环境持续交互
• 学习定价、库存、调度等决策策略
• 具备长程规划能力
• 基于真实反馈持续自我进化
• 根据企业偏好动态调整行为

这是强化学习 + 大模型 + 多智能体协同在真实商业系统中的落地，而不是模拟世界里的 benchmark。

⸻

工作重点（Focus）

• 为产业级 AI Agent 构建环境交互系统（业务状态 / 行为空间 / 奖励建模）
• 将强化学习引入真实场景，如渠道定价优化、库存分配、履约调度
• 构建 Agent 长程规划与复杂任务拆解能力
• 实现偏好学习与反馈优化（企业目标、风险约束、利润权衡）
• 设计仿真环境与离线评估体系，用于训练与回放真实业务策略
• 构建 Agent 学习闭环：感知 → 决策 → 执行 → 反馈 → 进化
• 搭建自动化训练、评估与部署流水线
• 提升大规模 RL 任务的稳定性与可观测性
• 重构 Agent / 数据 / 训练框架，使研究成果快速进入生产系统

⸻

理想经验

• 有强化学习 / Agent / 决策系统经验
• 能将现实问题抽象为状态、动作、奖励
• 熟悉 PyTorch / Python
• 有系统工程意识
• 对真实商业系统有好奇心

加分项：

• 多智能体系统
• 运筹优化 / 博弈
• 供应链 / 定价 / 资源调度经验
• LLM Agent 框架

⸻

你将面对的典型问题

• 一个渠道价格策略在不同区域表现完全不同，Agent 如何通过强化学习自动适配？
• 库存与履约冲突时，如何让 Agent 学会权衡利润、时效与风险？
• 新增一个业务模块，如何快速构建可学习环境？
• 真实业务数据噪声极大，如何设计鲁棒 reward？
• 企业偏好变化时，如何让 Agent 快速重新对齐目标？

⸻

技术栈

Python / PyTorch
分布式 RL
Agent 框架
Typescript / React（内部工具）

Department:
AI Agent Research Center
Location:
Hong Kong, Shen Zhen
Work Experience:
Graduate / Early Career
Number of openings: 10

About the Role

We are looking for Agent Reinforcement Learning Engineers to join our Agent Core team.

You will help build learning-capable AI agents that:

• Interact continuously with real-world business environments
• Learn decision policies for pricing, inventory, and operations
• Perform long-horizon reasoning and planning
• Optimize behavior through feedback and preference alignment
• Improve themselves over time in production

This role focuses on applying reinforcement learning together with large language models and agent architectures in real industrial systems — not simulated toy environments.

⸻

Focus

• Design agent-environment interaction systems (observations, actions, rewards)
• Apply reinforcement learning to real scenarios such as pricing optimization, inventory allocation, and fulfillment scheduling
• Build long-horizon planning and multi-step reasoning pipelines for agents
• Implement preference learning and feedback optimization (RLHF / RLAIF / online learning)
• Construct simulation environments and offline evaluation pipelines from real business data
• Build closed learning loops: Sense → Decide → Act → Feedback → Improve
• Develop automated training, evaluation, and deployment workflows
• Improve observability and stability of large-scale RL jobs
• Refactor agent, data, and training frameworks for production readiness

⸻

Ideal Experience

• Background in reinforcement learning, agents, or decision systems
• Strong Python + PyTorch
• Ability to abstract real-world problems into states, actions, and rewards
• Systems thinking mindset

Nice to have:

• Multi-agent experience
• Operations research / game theory
• Supply chain, pricing, or resource optimization exposure
• LLM agent frameworks (LangGraph, AutoGen, CrewAI)

⸻

Typical Problems You’ll Work On

• A pricing strategy behaves differently across regions — how should the agent adapt via reinforcement learning?
• Inventory and fulfillment objectives conflict — how can agents learn trade-offs between profit, cost, and service level?
• Business data is noisy and delayed — how do we design robust reward functions?
• Enterprise preferences shift — how do we quickly realign agent behavior?

⸻

Tech Stack

Python / PyTorch
Distributed RL
Agent frameworks
TypeScript / React (internal tools