Trajectory raises $45M and launches self-service model training

On November 2, 2026, Trajectory announced a $45 million Series A led by Sequoia at a $300 million valuation, with Nvidia, Bessemer and Bain Capital Ventures participating. It launched a self-service platform that brings model evaluation, training, sandboxes and deployment into one post-training workflow. The company’s larger aim is continual learning from production work, but two components intended to extend that workflow, Continual Grading and Group Size 1, are research previews.

In its November 2 report, TechCrunch described the 18-person startup as making its platform publicly available, named Harvey, Rogo, Decagon, Mercor and Nvidia as customers, and quoted co-founder Ronak Malde saying users decide “what good looks like for their organization.” For a newly interested AI team, the distinction is between the available self-service route for training against its own tasks and the production-data route, which is still in private beta.

The round backs a product built around expert judgment

Trajectory’s founders are Malde, Michael Elabd and Arjun Karanam, former Stanford classmates with research experience at Google DeepMind and Apple. Malde and Elabd worked at DeepMind; Karanam worked on generative foundation models at Apple. Their product is designed to let people who know a field’s standards direct post-training without building a model-research team and its infrastructure inside their own company.

Nvidia appears on both the investor and customer lists, giving the launch a named user alongside its venture backers. The financing establishes what investors committed and the valuation attached to the round; it does not measure revenue or show how well an unfamiliar customer will fare with the product. The commercial test is whether specialists at other organizations can turn their judgments into improvements that hold up on work the model has not seen during training.

The self-service route still needs a task and a grader

Trajectory’s training workflow separates benchmark evaluation from the tasks used to improve a model.

Trajectory’s developer documentation describes a benchmark route in which a team supplies tasks, an agent harness and a grader that scores outcomes. The team evaluates a base model, trains on its tasks, compares the resulting checkpoint on held-out tasks and selects a model to deploy. The same documentation labels the route that learns from production traces private beta. Public launch of the platform therefore does not mean every advertised way of improving a model is available through the same self-service path.

The division also identifies the work the buyer must do. A useful training task has to represent the behavior the buyer wants; its grader has to reward that behavior; and held-out work has to reveal whether the trained model improved beyond the examples it practiced on. Trajectory supplies the training and deployment machinery described in its materials, while the organization supplies the domain judgment that tells that machinery what to optimize. A convenient training run cannot correct a scoring rule that rewards the wrong answer.

How production work could feed the next model

The proposed continual-learning loop begins with traces of what an application’s agent attempted and what happened. Scores or expert judgments can turn those traces into feedback for training, after which a candidate model is evaluated before deployment. That sequence matters because a live interaction alone does not update a model: someone must capture the work, define a useful learning signal and decide whether the resulting checkpoint is better on separate tasks.

Continual Grading is intended to extend a small set of expert judgments into criteria for grading more production examples. Group Size 1 addresses a different obstacle: training from individual real-world interactions without reconstructing each task in a simulation. Both are presented as research previews, so the documented benchmark workflow and the fuller ambition of learning routinely from production traffic have different availability and maturity. The human decisions about what deserves a high score remain central even if the mechanics of training become easier to operate.

Customers show adoption, while results remain vendor claims

Trajectory documents a benchmark training route while marking training from production data as private beta.

Harvey and Rogo illustrate the specialist work Trajectory is targeting: legal and financial tasks where an organization may have its own definition of a satisfactory result. The broader customer list establishes named users, but it does not show that each uses the same features or has achieved comparable gains. Trajectory has described work with Harvey and Nvidia on legal-model training; those performance claims concern a particular model and evaluation setup, not a measured outcome for every team that starts using the platform.

The immediate product consequence is narrower and more useful than a promise that any deployed model will improve by itself. Teams with tasks, scoring functions and a way to evaluate held-out work have a documented route through evaluation, training and deployment. Whether that route can extend to their own live application traces depends on access to the private-beta workflow; broader access and results reported by customers will show how far the launch’s continual-learning ambition has reached.