Handshake, the college careers network that now supplies human-generated training data to AI labs, acquired Cleanlab, a startup that builds software to find and fix errors in AI datasets. TechCrunch reported the deal on January 28, 2026, the same day Cleanlab chief executive Curtis Northcutt announced it in a letter on the company's blog.
Terms were not disclosed. TechCrunch described the deal as mainly an acquihire: nine Cleanlab employees, including all three co-founders, are joining Handshake's research organization.
Handshake is also taking over Cleanlab's research and technology in confident learning, data-centric AI and the evaluation of large language models. Cleanlab's open-source data-centric AI package will stay open source and move to the more permissive Apache-2.0 license.
What Cleanlab built
Cleanlab was founded in 2021 by Northcutt, Jonas Mueller and Anish Athalye, who all hold computer science PhDs from MIT. Its algorithms flag mislabeled or unreliable data automatically, which can remove the need for a second human reviewer to check every label. It later added tools to evaluate the outputs of large language models.
Its open-source package has been downloaded more than one million times. Cleanlab says it is used by more than 100 large enterprises, and the company was named to the Forbes AI 50 list. TechCrunch reported that Cleanlab had more than 30 employees at its peak.
Cleanlab's approach targets the problem summed up by the phrase garbage in, garbage out: models trained or evaluated on bad data produce unreliable results. It describes its mission as making AI work reliably, and positioned its software upstream of model training, at the point where data is collected and labeled.
“If you're going to pick one, you should probably pick the source, not the middleman.”
Why Handshake is buying
Handshake started in 2013 as a recruiting platform for new graduates. About a year before the deal it launched a business that supplies expert human data, such as work by doctors, lawyers and scientists, to companies training frontier AI models. TechCrunch reported that it has worked with eight leading AI labs, including OpenAI, and was forecast to end 2025 with $300 million in annualized revenue. It was last valued at $3.3 billion in 2022.
Data quality is central to that business. Sahil Bhaiwala, Handshake's chief strategy and innovation officer, told TechCrunch that the company's in-house research team spends a lot of time studying where models are weak, and that Cleanlab's team had worked on the same problem for years.
In his letter, Northcutt said the Cleanlab team would lead AI research at Handshake with a focus on producing high-quality training data for frontier models.
Why Cleanlab chose Handshake
Northcutt told TechCrunch that other AI data-labeling companies had shown interest in buying Cleanlab, though he did not name them. He said rivals such as Mercor and Scale AI often use Handshake's network to find human experts, and argued that it made more sense to join the source of that talent than a middleman.
Cleanlab said Handshake's network includes about 20 million pre-vetted fellows, and that Handshake's AI data business reached a $100 million run rate within seven months. Northcutt described the combined company as pairing Handshake's talent network with Cleanlab's research arm to create a more integrated data lab.
The founders and the funding
Cleanlab raised $30 million in total. In 2023 it announced a $5 million seed round led by Bain Capital Ventures, followed by a $25 million Series A co-led by Menlo Ventures and TQ Ventures, with Bain Capital Ventures and new investor Databricks Ventures also taking part. Samsung Ventures and several angel investors also backed the company.
Northcutt thanked Aaref Hilaly and Matt Murphy of Menlo Ventures in his letter announcing the sale.
Cleanlab did not say what will happen to its commercial products for enterprise customers beyond the open-source package, which remains available under the new license.
The deal shows how human data suppliers are trying to move up the value chain. Labs paying for expert annotations increasingly want evidence that the data is accurate, and buying a team that has spent years building automated checks for label errors gives Handshake a research group focused on exactly that question.
