代理系统工程师
Agentic Systems Engineer
关于公司
Kantiv（Joist AI）是一家科技公司，正在改变建筑、工程和施工（AEC）行业专业人士管理营销和营收运营的方式。我们的AI驱动软件简化了工作流程，让团队更容易协作、创新并取得成功。
关于职位
Kantiv正在寻找一位拥有2-4年经验的工程师，帮助构建下一代代理应用，为AEC行业提供更高效的提案撰写体验。这些系统能够推理、使用工具、记忆并与用户协作。技术栈涵盖多代理编排、MCP服务器、技能、长期记忆、评估、检索以及使所有内容在生产环境中稳定运行的基础架构。我们希望找到一位能帮助我们构建这些系统的工程师。
你将负责
- 将代理构建为模块化、即插即用的组件，无缝融入整体技术栈。
- 在运行中的系统中添加记忆层（短期、长期、摘要、检索支持）。
- 集成工具、MCP服务器和技能。
- 负责你所发布的功能质量：测试、评估、可观测性等。
- 分析生产环境的追踪日志，了解系统实际行为，并通过修复关闭闭环。
- 我们期望的背景
- 2-4年编写生产级软件的经验。
- 强大的Python技能。你能写出高质量的Python代码，并能区分好代码和坏代码，尤其是在大量代码由LLM生成的当下。关注职责分离、良好的面向对象编程、符合习惯的语法、结构良好的模块，以及真正有效的测试。
- 对核心代理和LLM概念有扎实的理解：RAG、提示模式、工具使用、结构化输出、流式处理、上下文管理、基本的生成式AI原理。
- 你曾使用现代代理工具包构建过非平凡的项目，无论是个人项目、工作中原型，还是一个失控的黑客马拉松作品。
- 能够快速进入一个陌生的代码库并熟悉其结构。
- 对细节有敏锐的洞察力。你会在寻求解决方案前仔细思考问题。不会因为某个修复方案听起来很聪明就直接上手。在动手之前，你清楚到底哪里出了问题。
- 默认以数据为导向。决策来自生产追踪、评估数据和日志，而不是感觉。能够高效地分析追踪数据，找到真正的信号。
- 有Langfuse或LangSmith（或类似LLM系统的追踪/可观测性工具）的实际操作经验。
- 对前沿技术有真正的兴趣。
查看英文原文
About the company
Kantiv (Joist AI) is a technology company revolutionizing the way professionals in the architecture, engineering, and construction (AEC) industry manage marketing and revenue operations. Our AI-powered software streamlines workflows, making it easier for teams to collaborate, innovate, and succeed.
About the role
Kantiv is looking for an engineer with 2–4 years of experience to help build the next generation of agentic applications that streamline proposal writing for the AEC industry. These are systems that reason, use tools, remember, and collaborate with users. The stack spans multi-agent orchestration, MCP servers, skills, long-term memory, evals, retrieval, and the plumbing that makes all of it hold up in production. We're looking for someone to help us build it.
What you'll do
- Build agents as modular, plug-and-play components that slot cleanly into the wider stack.
- Add memory layers (short-term, long-term, summarization, retrieval-backed) into running systems.
- Wire up tool integrations, MCP servers, and skills.
- Own quality of the features you put out: tests, evals, observability, the works.
- Dig into production traces to understand what the system is actually doing, and close the loop with fixes.
Background we're looking for
- 2–4 years of writing production software.
- Strong Python skills. You write good Python and can tell good Python from bad, especially now that a lot of code comes out of an LLM. Separation of concerns, clean OOP, idiomatic syntax, well-structured modules, tests that actually test something.
- Solid grounding in core agentic and LLM concepts: RAG, prompting patterns, tool use, structured outputs, streaming, context management, basic generative AI fundamentals.
- You've built something non-trivial with the modern agent toolkit, whether that's a side project, a prototype at work, or a hackathon thing that got out of hand.
- Able to drop into an unfamiliar codebase and find your way around fast.
- A keen eye for detail. You sit with a problem before reaching for a solution. No jumping to the shiny fix because it sounds clever. You understand what's actually broken before you touch anything.
- Data-driven by default. Decisions come from production traces, eval numbers, and logs, not vibes. Comfortable slicing through trace data to find the real signal.
- Hands-on experience with Langfuse or LangSmith (or equivalent tracing/observability for LLM systems).
- Genuine curiosity about the frontier. You read the blog posts, try the frameworks, and have opinions about where agent design is headed.
Experience we'd be particularly excited about
- Search and retrieval: embeddings, vector databases, hybrid retrieval, rerankers, and the gap between a retrieval system that demos well and one that survives real data.
- LLM evaluations end-to-end: designing evals, choosing what to measure, building the harness, keeping scores honest as models and prompts shift.
- LangGraph depth: building custom graphs, understanding checkpointers, working with context-management nodes (summarizers, windowing, state pruning) inside larger agent graphs.
What to expect
We conduct a rigorous interview process based on integrity, talent, and drive. We trust our teammates from day one and move quickly to evaluate your fit for the role. The entire interview process typically takes two weeks. Here's what to expect:
- A 30 minute Zoom meeting to talk about Kantiv, your background, and answer any questions about the role. (Getting to know each other)
- 45-minute Python proficiency / agentic coding proficiency test. 2 problems. 1 to be coded by hand. Other using Gen AI.
- 60 min project, deep dive into the work they have done. A short presentation followed by a Q&A. Presentation should conclude between 20-25 min.
- 45 min interview on Gen AI / LLM fundamentals.
- 30 min culture fit.
Originally posted on Himalayas