Trajectory raises $45 million to help specialists train AI through feedback

Trajectory has secured $45 million to tackle a costly step between a general AI model and one that knows a business’s work. Sequoia led the Series A at a $300 million valuation, backing a platform designed to let specialists guide model improvements without assembling their own AI research teams.

Nvidia, Bessemer Venture Partners and Bain Capital Ventures also participated. The financing backs an 18-person startup founded about a year ago by three Stanford classmates whose experience includes Google DeepMind and Apple.

From a general model to experience on the job

Trajectory’s starting problem is that deployed AI models do not automatically learn from their mistakes. Businesses adapting them to law, finance or other specialized work often need further training. That can mean hiring an internal research team or paying outside consultants known as forward-deployed engineers to help tailor the technology.

Co-founder Ronak Malde described the gap to TechCrunch by comparing a foundational model to a highly capable PhD student who is always beginning a new job. His argument is that general intelligence is not the same as the accumulated experience a business needs for its particular tasks.

But what the economy values is someone who worked on the job for 30 years

Ronak Malde, Trajectory co-founder, speaking to TechCrunch

Turning feedback into another training run

The company’s response is a platform for continual learning: repeatedly using feedback to retrain a model for specialized work. A business starts with an open-source model and connects Trajectory to its applications. The platform collects feedback about performance, then uses that information for further training intended to improve subsequent results.

That work is called post-training—additional training after a model’s initial development. Malde says lawyers, doctors and other specialists can use Trajectory to improve models almost automatically. They still decide what counts as good performance for their organization; the platform is meant to turn those judgments into changes to the model.

Harvey and Rogo already use the platform to post-train open-source models for legal and financial work. The tasks are narrower than simply answering questions about those fields: they include reviewing a particular kind of contract or valuing a company with a specific methodology. Decagon, Mercor and Nvidia are also customers, TechCrunch reported.

Automation with people still setting the standard

The founders put human involvement at the center of their safety argument. Co-founder Michael Elabd says keeping people in the training process makes Trajectory’s approach safer than autonomous self-improvement, in which AI systems improve their own capabilities. He argues that the latter could produce unintended or dangerous outcomes. That is the company’s position, not an independently demonstrated safety result.

Trajectory is competing with other businesses trying to automate post-training. It identifies Thinking Machines Lab, the company behind the fine-tuning platform Tinker, as a direct rival, alongside Applied Compute and Prime Intellect. Malde points to Trajectory’s team as its main advantage rather than claiming the company has the field to itself.

Malde and Elabd previously worked at DeepMind, while co-founder Arjun Karanam worked on generative foundation models at Apple. The startup has also recruited researchers and product designers from OpenAI, Stripe and Netflix. Karanam says its interface makes post-training more accessible to nontechnical customers—a central part of its effort to let domain experts guide training themselves.