Trajectory raises $45M to make AI model training accessible to nontechnical experts

Trajectory, an 18-person startup founded about a year ago, raised $45 million in a Series A at a $300 million valuation. Sequoia led the round, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures joining.

The problem it's going after is real. Today's models don't get better at a job from their own mistakes unless someone retrains them, and that usually means hiring researchers or expensive forward-deployed engineers. Trajectory wants to automate the loop: connect an open-source model to a company's app, gather feedback on how it performs, and retrain it on that feedback over time.

Harvey and Rogo already use the platform to post-train open models for legal and financial work, and the company is now opening it to the public.

We like the shape of this. The most valuable knowledge in a firm sits with the people who do the work, not with the people who tune models. A tool that lets a senior lawyer or analyst define what good output looks like, and have the model actually learn it, puts that judgment where it belongs. The co-founder's line about models being like a bright new hire who never gets to build up years on the job is a fair description of the gap.

The caution is in the word automatic. A model that keeps learning from feedback is only as good as that feedback. If reviewers are rushed, inconsistent, or wrong, the model learns the wrong lessons quietly, and in law or medicine those errors carry real costs. Continual learning also makes the system a moving target. The model that passed review in spring may not be the same one answering questions in fall. We want to see versioning, rollback, and audit trails treated as features, not afterthoughts.

The company makes a point of separating itself from recursive self-improvement, arguing that keeping humans in the training loop is safer. On the same day Anthropic's filing put that topic in the headlines, that's a smart distinction, and we think it's mostly right. Humans setting the target is a real check. It's only a check, though, if those humans have time to look closely.

The market is crowded. Thinking Machines Lab, Applied Compute and Prime Intellect are all going after post-training. Trajectory's pitch is its team and an interface built for people who don't write code. Design really can be the moat here: if domain experts find it confusing, the feedback dries up and the whole loop stalls.

The three co-founders met at Stanford. Ronak Malde was among the researchers Google DeepMind acqui-hired from Windsurf, Michael Elabd also worked at DeepMind, and Arjun Karanam worked on generative foundation models at Apple. Other listed customers include Decagon, Mercor and Nvidia.