Trajectory, a startup that wants subject-matter experts to improve AI models without writing code, has raised a $45 million Series A. The round values the company at $300 million and was led by Sequoia, according to TechCrunch and Wowtale. Nvidia, Bessemer Venture Partners and Bain Capital Ventures also took part.
The idea. General-purpose models lack the value of someone who has worked on a job for 30 years, said co-founder Ronak Malde, according to TechCrunch. Trajectory builds an interface so that lawyers, doctors and other specialists can post-train models, meaning refine them on their own data and examples. The company says specialists can improve models nearly automatically.
The company. Trajectory has 18 employees and was founded about a year ago by three Stanford classmates, Malde, Michael Elabd and Arjun Karanam, who previously worked at Google DeepMind and Apple. Malde was among researchers Google DeepMind hired from Windsurf. Other coverage says the company raised a $15 million seed round in May, bringing total funding to about $60 million.
The customers. Harvey and Rogo use the platform to post-train open-source models, and Decagon, Mercor and Nvidia are also customers, according to TechCrunch. The post-training platform is now publicly available. Another report describes the platform as self-service, while noting that a workflow that uses a company’s production data remains in private beta.
The competition. TechCrunch names Thinking Machines Lab, which sells a tool called Tinker, Applied Compute, valued at $3.25 billion last month, and Prime Intellect as rivals. Trajectory has recruited researchers and product designers from OpenAI, Stripe and Netflix.
The safety argument. Elabd argues that keeping humans involved makes the approach safer than autonomous self-improvement by AI, according to TechCrunch. Karanam, who worked on generative foundation models at Apple, said the company built an interface for non-technical customers.
What is not stated. The article does not give revenue, pricing, how the money will be used, or the number of customers beyond those named. Performance claims in other coverage relate to specific models and tests, not to a measured outcome for every team, so they should be checked.
Why it matters. Fine-tuning models usually needs machine-learning engineers. A tool that lets domain experts do it could spread AI into professional work faster, and put more control in the hands of people who know the field. It also fits the trend of companies customizing open models rather than relying only on general-purpose chatbots.
Sources: TechCrunch and Wowtale. Details are as reported by those outlets and the company.
San Francisco skyline from the Embarcadero (representative image), Wikimedia Commons, CC0
