AI Engineer (Agent Platform)

About the role

You’ll build the AI that does the actual work in people’s documents. Our agent today helps users inside the product; the next step is bigger — an agent that takes a document task end-to-end from wherever the user is: inside Lumin, in Slack, in email. The bar for all of them is simple: give it a task, walk away, and come back to finished work you can trust. You’ll design how the agent plans and recovers, which tools it uses, and how we prove it’s getting better instead of just hoping it is.

What your first six months look like

Ship a first version of the agent that completes document tasks on its own — not perfect, but good enough that users review the result instead of watching every step

Build the eval harness that makes changes safe to ship: prompt and model changes — whether the changes made by human or an AI agent — must prove they make things better before anything goes out

Responsibilities

◦ Design and build agent workflows in Python (we use LangGraph): planning, state management, tool use, recovery when things go wrong.
◦ Define the agent’s capabilities as clean, modular tools — the interface other engineers build against.
◦ Own reliability: guardrails against prompt injection, tools that respect user permissions, human approval for destructive actions.
◦ Build and run evals — golden datasets, regression gates, latency budgets — and use them to gate every change.Route models across vendors to hit quality and latency targets per capability.
◦ Instrument everything: tool usage, failure modes, costs — use traces to find the next fix.
◦ Partner with the product team that owns the UI, so agent features appear where users already work.

Your skills and experience

◦ 2+ years as a software engineer, AI/ML engineer, data engineer, or equivalent with production Python.
◦ Hands-on experience building LLM applications: tool calling, structured outputs, prompt engineering, agent loops (we use LangGraph; similar frameworks are fine).
◦ Eval mindset: you think in test sets and want a measurable definition of “better”, even for fuzzy outputs.
◦ Production fundamentals: async Python, message queues, APIs (gRPC/REST), logging you can debug from.
◦ Comfortable working in English, spoken and written.

NICE TO HAVE

◦ An agent you shipped to real users, especially one connected to a third-party platform (Slack, Teams).
◦ LangGraph and Langfuse specifically.
◦ Classical ML or fine-tuning background.