Job description

Type: Freelance / Contract

Location: Fully Remote

Hours: Around 20 hours/week, with flexible scheduling

Process: Short screening call → Technical assessment → Onboarding

The work

You'll build realistic machine learning engineering problems that AI models learn from. Each one is a small but genuine codebase with something wrong in it, packaged in Docker, plus an automated grader that checks whether the model really solved it. Examples of what you might build:

A model that looks excellent in testing but falls apart in production, because one of its features quietly gives away the answer
An evaluation that reports misleading numbers because of a subtle data error
A sluggish data pipeline that has to run much faster without changing a single result

The most interesting part is the grading. Any correct fix has to pass, however it's written, while shortcuts, hardcoded answers and faked results have to fail.

What you'll need

A degree in computer science, engineering, maths or a related subject
At least 4 years of full-time experience in ML engineering, applied ML or data science engineering
Strong production Python, including pandas, NumPy, scikit-learn and at least one of XGBoost, LightGBM or CatBoost
Confidence with Docker, and with finding your way around large codebases spread across many files
An instinct for spotting leakage, flattering metrics and solutions that only look right

Previous AI-training work is a plus, not a requirement. If you've spent years catching the bugs other people missed, you'll probably be good at this.

What's on offer

Between $30 and $100 per hour, depending on experience
Fully remote, with hours that fit around you
Technically demanding work at the sharp end of AI development

How to apply

Send us your CV. Shortlisted candidates are invited to a 30-minute technical interview, where you'll review an example task and talk through how you'd design one.