By Mufy Pachorawala, founder of getAIwork · Last updated: September 3, 2026
Key facts
- Founded 2024 by Kat Hu and Robi Lin, both previously at Turing, and part of Y Combinator’s Summer 2024 batch
- Acquired by Mercor.io Corporation on February 6, 2026, per the deal announcement from Mercor’s legal advisers Orrick
- Network of 20,000+ experts across STEM and professional services, including academic PhDs, medical professionals, business analysts and finance consultants (company figure)
- What it actually sold: training data for fine-tuning and RLHF, frontier benchmarking, safety evaluations and RL environments
- Team of roughly 15 at the time of acquisition, based in San Francisco
- 591 AI platform programmes are live on the free getAIwork board as of September 3, 2026, screened and human-approved
What was Sepal AI?
Sepal AI was a data-development company for advanced AI. It built what the industry calls golden datasets: small, expensive, expert-authored collections used to measure how good a model really is, rather than the large cheap piles used to make it fluent in the first place. Alongside that it produced evaluation benchmarks, safety evaluations run before deployment, and reinforcement learning environments.
The company was founded in 2024 by Kat Hu and Robi Lin, who had both worked at Turing on large language model training infrastructure, Hu on go-to-market and operations and Lin on product and fulfilment. It went through Y Combinator’s Summer 2024 batch and stayed small, around fifteen employees in San Francisco, while the expert network attached to it grew past 20,000 people.
That ratio is the whole business model in one line. Fifteen staff, twenty thousand contributors. Sepal was not trying to employ experts, it was trying to reach them: cell biologists to write benchmark questions about cell biology, finance consultants to grade finance answers, doctors to check medical reasoning. The staff built the pipelines. The network supplied the judgement.
Sepal AI was acquired by Mercor in February 2026
On February 6, 2026, Mercor.io Corporation announced it had acquired Sepal AI. Mercor’s stated reason was to strengthen work on long-horizon research and on training models against the most complex tasks humans perform, and it described Sepal as a team working at the intersection of human data, reinforcement learning environments and specialised research projects.
This matters more to a would-be contributor than it sounds. Acquisitions in this category are usually about the network, not the software. Mercor already runs one of the largest expert marketplaces in AI work, and Sepal’s twenty thousand credentialled experts were the obvious asset. The two networks overlap in exactly the way that makes a merger tidy and an outside observer confused about which brand to apply to.
| Before February 2026 | After the acquisition |
|---|---|
| Sepal AI recruited experts directly for benchmark and RL environment projects | That expert demand is now channelled through Mercor’s network and intake |
| Two separate application routes for people with the same credentials | One route, with Mercor’s own screening and matching in front of it |
| Sepal published its own project types: cell biology benchmarking, finance question and answer evaluation, human baseline testing | The same categories of project continue, under Mercor’s contracts and payment terms |
| Small company, fifteen staff, limited support capacity | Larger organisation with published payment and onboarding processes |
Sepal is one thread in a category with 591 live platform programmes on the getAIwork board. Two minutes matches you against all of them.
Take the 2-minute match quiz →
Can you still apply to Sepal AI?
Not as a standalone platform in the way you could in 2025. The company was acquired in February 2026 and its expert-facing intake is no longer being marketed as a separate product. Searches for a Sepal AI application form now mostly surface third-party listicles that were written before the acquisition and never updated, which is a good general warning about this whole category of content.
If a page is currently asking you to pay a fee, buy a course, or hand over bank details to apply to Sepal AI, treat it as an impersonation. Legitimate expert platforms in this market earn from the client side. None of the established ones charge the worker to apply, at any stage, and that single rule filters out most of the fraud aimed at this audience.
What did the work actually involve?
Sepal’s published project examples give an unusually clear picture of what expert AI data work looks like when it is aimed at specialists rather than at general annotators. Three named examples were cell biology benchmarking, finance question and answer evaluation, and human baseline testing.
Human baseline testing is the one most people have not heard of and the one that explains the pay in this category. To claim a model performs at expert level on a task, somebody has to establish what expert level actually is, by having real experts attempt the same task under the same conditions. You are not labelling data. You are being the control group. That requires the credential, and it is why work of this kind is priced against professional consulting rather than against task work.
| Project type | What the expert does | What it needs |
|---|---|---|
| Benchmark authoring | Write questions in your field that a strong model should get wrong, with defensible correct answers | Deep subject knowledge, usually postgraduate or professional |
| Evaluation and grading | Score model answers against a rubric and explain the reasoning behind the score | Judgement plus the ability to write clearly under a rubric |
| Human baseline testing | Attempt the same tasks the model is being tested on, so a comparison exists | The credential itself, and honest effort |
| RL environment design | Build a task environment with a checkable success condition a model can be trained against | Subject knowledge plus some technical comfort |
None of that is beginner work, and it is worth saying plainly rather than dressing it up. On our own board, only 162 of 1,435 live listings are tagged beginner-friendly. Specialist evaluation work sits at the opposite end of that distribution.
Who this kind of work suits, and who it does not
It suits people who already hold a credential they earned for another reason. A radiologist, a tax accountant, a materials scientist, a litigator, a translator with a rare pair. The credential is the product. The AI part is incidental, and platforms in this category are quite open about the fact that they want your professional judgement rather than any AI skill.
It does not suit someone looking for steady predictable hours. Project-based expert work arrives in bursts tied to whichever lab is currently buying, goes quiet without warning, and offers no queue position while you wait. That is not a complaint about any one company, it is the structure of the whole category, and anyone telling you otherwise has not done it.
Where this work lives now
The practical answer for a specialist in 2026 is to apply to two or three of the larger expert networks rather than chasing a brand that has been absorbed. Our Mercor review covers the acquirer directly, and our comparison of Outlier alternatives covers the wider field including the platforms that recruit for evaluation rather than annotation.
One practical note that applies across all of them: these platforms are non-exclusive and free to join, so applying to several in parallel costs nothing but evenings. What it buys you is exposure to more than one demand cycle, which is the only real protection against the quiet stretches.
1,435 live AI listings, 591 of them platform programmes, screened and human-approved. Filled listings are deleted daily.
Frequently asked questions
Is Sepal AI legit?
Yes. Sepal AI was a real Y Combinator Summer 2024 company founded by two former Turing staff, with named investors and frontier AI labs as clients. It was acquired by Mercor on February 6, 2026, which is why its own sign-up route is no longer active.
Who acquired Sepal AI?
Mercor.io Corporation, announced on February 6, 2026. Mercor said the combination would strengthen its work on long-horizon research and on training models against complex human tasks.
Can I still sign up to Sepal AI as an expert?
Not as a standalone platform. Since the February 2026 acquisition the expert demand runs through Mercor. Any page currently charging a fee to apply to Sepal AI should be treated as an impersonation, because legitimate platforms in this category charge clients, never workers.
What did Sepal AI pay experts?
Sepal never published a public rate card, so we do not quote one. Rates in expert evaluation work are set per project and per credential rather than by a single platform-wide figure, and any site quoting you a guaranteed Sepal AI rate is inventing it.
What kind of experts did Sepal AI recruit?
STEM and professional services: academic PhDs, medical professionals, business analysts and finance consultants, across a stated network of more than 20,000 people. The common thread was a credential earned outside AI.
What is a golden dataset?
A small, expert-authored dataset used to measure model quality rather than to train general fluency. Because every item has to be right, it is written by credentialled people and costs far more per item than bulk annotation.
Is expert AI evaluation work beginner-friendly?
Generally no. It is bought for the professional credential behind the judgement. On the getAIwork board only about 11 percent of live listings are tagged beginner-friendly, and specialist evaluation work is not among them.
What should I do instead of applying to Sepal AI?
Apply to the larger expert networks directly, including Mercor, which now holds this demand. Applying to two or three in parallel is normal, costs nothing, and exposes you to more than one project cycle.
Mufy Pachorawala · Founder, getAIwork
My AI scans thousands of AI-job posts (39,847 screened so far) and I personally approve every listing before it reaches the board. I write these articles by the same rules: pay quoted only as listed, and no income promises. Read our editorial rules.
Related: Mercor review · Mercor vs Micro1 · 12 Outlier alternatives compared · AI training platforms hiring now · AI job market statistics
