What you would do

Write data-analysis and modeling tasks with one checkable answer and a planted pitfall, such as a feature that leaks the label.
Grade model answers on experiment design, metrics and statistical validity, and mark the first step that goes wrong.
Review training and evaluation code, such as a PyTorch training loop or a scikit-learn pipeline, for bugs that quietly change the results.
Design rubrics for grading model outputs, so different graders reach the same score.
Write reference notebooks that show the right analysis step by step.
Compare two model explanations of an ML idea, such as regularization or attention, and say which is right and why.

What we look for

You study at or graduated from a college or university in a quantitative field, such as computer science, statistics, math or engineering.
One substantial ML or data project you did yourself, with or without AI tools, that you can explain in depth, including how you checked the results.
You read and write Python with common data and ML libraries, such as pandas, NumPy, scikit-learn or PyTorch.
You check for leakage, bad splits and misleading metrics before you trust a number.

A degree in computer science, statistics or a related field; a master’s or PhD is even better.
Research or production ML work, especially at an AI lab or AI company.
Hands-on work with language models: fine-tuning, evaluation, red-teaming or building graders.
Papers, competition results, models or datasets others use.

Task size: Tasks take about 30 minutes to 5 hours. Each one shows its estimated time and price before you start, and you are paid that price even if it takes you more or less time.
Pay: Per task accepted in review: its estimated time times the project’s rate. A task estimated at 2 hours at $150 an hour pays $300. Accepted tasks are paid within 30 days.
Hours: Projects last 1 month to a year. Each sets a weekly minimum, usually about 20 hours of estimated work, and you choose when you work.
Tools and review: Your own computer with Python and Jupyter, plus the libraries each task names. Every task is reviewed against the project’s written quality standard before it is accepted.