AI/ML Software Engineer-Task Creator (RL Environments) (Contract)

Careerflow Human Data Labs partners with AI companies to bring real-world professional expertise into their products.

About the work

We build hard, realistic tasks used to test how well AI agents do real work on a computer.

Each task is a small scenario running on a Linux desktop — real files, real applications, real documents — plus a program that automatically checks whatever the agent produced.

You create those tasks end to end.

The bar: hard for a top AI agent, straightforward for a competent human.

· Design a realistic multi-step scenario across a few applications.

· Put together the files it needs — spreadsheets, documents, emails, data.

Sometimes supplied to you, sometimes built by you.

· Write the instruction the AI agent receives: clear, complete, no giveaways.

· Write Python to set up the environment and to score the result automatically.

· Test it, run it, and hand over short documentation.

What we are looking for

· Python — comfortable writing and debugging real scripts.

· Heavy AI user — you work with AI coding tools and models every day and get genuinely good output from them.

· Some AI / ML background — worked on AI projects, at an AI company, or with agents and evaluations.

· Ubuntu and virtual machines — basic comfort.

Everything runs on hosted Linux VMs, so you should be fine on a terminal and inside a Linux desktop.

· Careful and detailed — an unclear instruction or a sloppy check breaks the task.

· Git basics.

A strong plus

· You have built RL environments before — anything where an agent acts and a program scores the outcome.

· You have built tasks or benchmarks for computer-use agents (agents that control a real desktop, browser, or operating system).

· You know OSWorld or similar computer-use benchmarks.

Not required, but it means you will be productive on day one.

Payment and availability

· $50 USD per task , paid once the task is accepted and approved in review.

· One free revision round if a task needs fixes.