Trajectory raises $45M to turn expert feedback into training

Trajectory has raised $45 million to make domain experts’ judgments usable as model-training signals, rather than leaving every company to assemble its own post-training stack. The startup announced a Sequoia-led Series A at a $300 million valuation on November 2, with participation from NVIDIA, Bessemer and Bain Capital Ventures.

Its platform launch targets companies that want specialized models to improve from production tasks and feedback. The company’s announcement describes a connected training, evaluation and deployment workflow: supply tasks and a way to score them, train a model against those standards, evaluate the result and decide whether to ship it.

TechCrunch reports that customers including Harvey and Rogo use Trajectory to post-train open-source models for legal and financial work. The idea is to capture what experienced specialists consider a good answer, then use that feedback in further training. A model seeing a correction in a conversation is not, by itself, the same as updating its weights; Trajectory’s pitch concerns the infrastructure connecting feedback to subsequent training.

Two routes toward learning from real work

The launch introduces two features explicitly labeled research previews. Continual Grading aims to extend a small set of expert judgments into grading criteria across more production data. Group Size 1 is presented as a way to train on individual production interactions without first constructing a large training environment.

Those are company-described capabilities and ambitions, not evidence that every organization’s model will improve automatically. The quality of task definitions, feedback, evaluations and release decisions still matters. Trajectory’s homepage says model updates pass through a customer’s evaluation suite and approval workflow before reaching production.

The platform is a hosted training service, not a new foundation model, a coding-agent plugin or proof of autonomous recursive self-improvement. TechCrunch describes keeping human specialists involved as part of the founders’ approach.

A launch with an access caveat

Trajectory’s announcement calls the product self-serve and says users do not need a sales call to start. At review time, however, the homepage’s main access button still led to a waitlist. The launch wording therefore should not be read as a verified promise that every visitor can immediately obtain an account and start a training run.

The reviewed launch materials do not establish a public price or universal savings figure. AITrending has not run a training job or independently measured the company’s performance claims. The concrete development is a newly financed platform built around a practical bottleneck: turning expert feedback from real work into a repeatable model-development process.