Trajectory bets the people who train AI won't be ML engineers

Trajectory is about a year old and has 18 employees. It has just 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 also participating.

The pitch is simple. Today's AI models don't learn from their own mistakes unless someone trains them again. Companies in fields like law and finance usually fix that by fine-tuning a general model. That takes an in-house research team or pricey outside consultants, the forward-deployed engineers of the trade.

Feedback in, better model out

Trajectory wants to automate that loop. A company starts with an open source model and connects Trajectory's platform to its applications. The platform collects feedback on how the model performs, then uses it to retrain the model. The approach is called continual learning, and the aim is steady improvement at narrow tasks.

The notable part is who steers it. Co-founder Ronak Malde says specialists such as lawyers and doctors can improve the models nearly automatically, by deciding “what good looks like for their organization.” The company says it built an interface to make post-training approachable for nontechnical users.

There is early traction. Harvey and Rogo use the platform to post-train open source models on work like reviewing a certain type of legal contract or valuing a company with a particular method. Decagon, Mercor and Nvidia are also customers. The platform is now open to the public.

What the round signals

The valuation is a bet that post-training becomes a product that domain experts operate, rather than a service sold by engineers. If that holds, the scarce skill in specialized AI shifts from knowing how to train a model to knowing what a good result looks like.

The field is crowded. Thinking Machines Lab, the company behind the fine-tuning platform Tinker, is in it. So are Applied Compute and Prime Intellect. Applied Compute was valued at $3.25 billion last month, more than ten times Trajectory's price tag. That suggests investors see the category as large, and that Trajectory still has a lot of ground to cover.

Co-founder Michael Elabd offers one differentiator. Trajectory's method resembles recursive self-improvement, where AI upgrades its own abilities, but it keeps humans in the loop. He argues that makes it safer than fully autonomous improvement. Whether customers weigh that as heavily as speed and cost is an open question.