1,230+ trainers in our network
Curatrain tests engineers, scientists, doctors, lawyers and linguists with assessments an AI can't sit for them, then matches them to AI companies' active, paying projects.
Specialists we vet and place
What we do
Every project is staffed with experts who passed our skill tests for the skills it needs.
DataExperts label text, images, audio and code, and write down why, so every label can be checked.
Fine-tuningSpecialists write the answers a model should learn from, rank its replies, write grading rubrics and try to break it.
Virtual environmentsEngineers and domain experts build realistic tasks, tools and checkers for AI agents, then judge how the agents did.
3D and gamesPeople who model in 3D and can write a game from scratch make and judge the visual, 3D and game data that multimodal models learn from.
From the first method to the newest. In every one, the model learns from work that a person who knows the subject has to do.
The expert: Writes the ideal answer to each prompt. The model learns to answer the same way.
OpenAI, InstructGPT · 2022 (opens the source)The expert: Compares two answers and picks the better one. A reward model learns from the choices.
Anthropic · 2022 (opens the source)The expert: Writes problems a program can check, so the model is rewarded only when it is right.
Ai2, Tülu 3 · 2024 (opens the source)The expert: Writes a reference answer and the checklist a good answer must meet, for a judge to score against.
Scale AI · 2025 (opens the source)The expert: Builds a realistic work task with apps and files, and the check that says whether an agent did it.
Epoch AI · 2026 (opens the source)How it works
One account takes you through every stage. Each step unlocks the next, and everything you earn stays on your profile.
Step 01You
Sign up with just your email, no password. Upload your resume and our AI fills in your profile. Nothing is saved until you confirm each field.
Fig. 01 · Resume check, live screen
Step 02You
Ten questions, one at a time, against the clock. Two ask you to explain your answer in your own words. Tests run in fullscreen with your camera on for a few still photos, never video.
Python test · Question 4 of 10
1:12 leftWhat does sorted(words, key=len) return when words = ["ab", "c", "def"]?
Fig. 02 · Skill test, example question
Step 03You
Each skill stays locked until you pass its test. You see your score when you finish, and a retake can only raise your level.
Skill cards
2 unlockedPython
SQL
Clinical reasoning
Contract law
Fig. 03 · Skill cards
Step 04Curatrain AI
Our AI looks at every active, paying client project, asks you a few questions, and matches you to the ones that need your skills.
Matched for you
Code review for a coding model
Strong fitSQL reasoning evaluations
Good fitHave you reviewed pull requests on a production codebase?
Fig. 04 · Matches
Step 05The client
The client sees your verified profile and skill levels through a secure link, and approves the match.
Review your match
Secure link
Ada O. Example
Machine learning engineer · 5 yrs
Fig. 05 · Client approval
Step 06Curatrain AI
Our AI negotiates your hourly rate and aims close to the client's budget, not the lowest rate you would accept.
Rate
ExampleThe client's budget allows $66 an hour for this project. Shall I confirm it?
Yes, confirm $66.
Rate agreed · $66 an hour
Fig. 06 · Rate agreement, example figures
Step 07You
You get a Curatrain workspace email and a Slack channel with the team, and your first tasks are waiting.
Day one
All setFig. 07 · Onboarding
For experts
AI trainers keep reporting the same problems: bans with no reason, unpaid tests followed by silence, and being vetted for work that never comes. Curatrain is built to answer each one.
We only match you to projects a client is running and paying for now. Nobody is vetted into a pool and left waiting.
If anything on your account is restricted, we tell you why, and a person reviews your appeal.
No unpaid test followed by weeks of silence. You see your score when you finish, every score is kept, and you get up to ten attempts per level.
86/100Example: Level 3 unlocked
Tests take a few still photos and run their face and phone checks on your computer. Nothing is recorded as video or sound, and your resume is downloadable by you alone.
Our AI negotiates for you and aims for what the client has budgeted, not the lowest rate you would take.
Remote, worldwide
Tests and projects are fully remote. Experts work from home, a café or a beach, in whatever time zone they live in.
Worldwide, Africa first
Much of the world's AI data work is done by people in Africa, usually through layers of vendors and agents. Curatrain works with experts everywhere and puts Africa first: you hire tested people directly.
Africa is full of brilliant, creative people: engineers, doctors, lawyers, linguists, artists and game makers. We test them properly and put them in front of the teams building AI, with nothing in between.
Usually
Each layer takes a share before the pay reaches the person doing the work.
With Curatrain
Experts are paid close to the client's budget, never under $7 an hour.
78
data labelling and moderation centres across 39 of Africa's 54 countries. Nigeria has the most.
Data4Mods map, Personaldata.IO and the African Content Moderators' Union$12.50 → $2
One AI company paid its vendor $12.50 per worker hour; the workers in Kenya took home $1.32 to $2. The vendor said $1.46 to $3.74.
TIME, January 2023Middlemen
Kenyan annotators training AI models told reporters they dealt only with agents and third-party firms, never the company.
Rest of World, December 2025For AI companies
Every Curatrain expert has a confirmed resume, graded skill levels and tests taken under anti-cheating and camera checks. You approve each match yourself.
We match from experts who have already passed the tests your project needs, so you review a few strong profiles.
Graded tests for each skill and level. You only see what an expert has proven, with the resume read and confirmed.
One person per account, bot checks at sign-in, behaviour checks during tests, and questions where they explain their answer, not just pick it.
Approve through a secure link. Minimal personal data, private file storage, and no AI lab owns Curatrain.
Shortlist · 1 of 3
Secure linkAda O. Example
Machine learning engineer · 5 years
Lagos · GMT+1 · 30 hours a week
The research
Ten studies, newest first. Models keep improving, and they still fall short where experts are needed to write the work, check it and judge it.
Nanyang Technological University
AI judges rated AI-written research questions as highly novel. Domain experts disagreed and preferred the questions from the human authors' papers.
UC Berkeley, with 250+ industry experts
On long, paid-work tasks designed with industry experts, frontier agents fully passed under 1% of the hardest tier.
Stanford HAI
Synthetic data is still not replacing real data in pre-training. The gains came from curating real data better.
OpenAI
On open-ended research tasks written and checked by PhD scientists, OpenAI's best model scored 25%.
Center for AI Safety and Scale AI, in Nature
Humanity's Last Exam: questions written by subject experts still beat frontier models, which are often wrong and sure of themselves.
Epoch AI
Builders of training environments named keeping expert-made tasks good at scale as their top bottleneck, and domain knowledge as worth more than ML skill.
Center for AI Safety and Scale AI
At launch, the best AI agent fully completed 2.5% of real freelance projects that professionals had already delivered.
FAIR at Meta and Virginia Tech
Training on AI-rephrased text alone was no faster than on human-written text. The best mixes stayed mostly human, about 30% synthetic.
Oxford, Cambridge and others, in Nature
Training on AI-generated content without care leaves defects in the next models that do not go away.
Epoch AI
Training runs will use up the stock of public human-written text some time between 2026 and 2032.
Summaries are ours. Numbers are as each source published them; some have changed since.
Specialists who can write and grade data for AI models: software engineers, data scientists, physicians and nurses, lawyers, linguists and translators, scientists, mathematicians and finance professionals.
No video. Your camera stays on during a test, but we only take a start photo and a few still photos, and the checks for phones and other people run on your own computer. Nothing is recorded as video or sound, a person reviews anything the camera flags, and we tell you exactly what is tracked before a test starts.
You can try again, up to ten attempts per level, and every score is kept. Nobody fails on a single signal.
Matching starts from clients' active, paying projects. Our AI only matches you to projects that need the skills you have unlocked, so you are never vetted for work that does not exist.
Our AI negotiates it with you and aims close to the client's budget for the project, not the lowest rate you would accept. You agree the rate before you start.
You get the reason in writing and can appeal. A person, not a bot, reviews every appeal.
Each expert's resume is read and confirmed, then they pass graded tests for every skill and level they claim, with anti-cheating checks and an AI interview that asks them to explain their own work.
One person per account, bot checks at sign-in, a flag when the same resume appears on two accounts, and behaviour checks while each test runs.
We send you each match through a secure link with the expert's verified profile and skill levels. You approve or pass.
Email hello@curatrain.com with the project, the skills it needs and your budget. We match from experts who have already passed those tests.
Experts start with an email address. Companies start with a conversation.
