Trajectory's new platform lets a company post-train its own model on the tasks it already runs and the feedback those runs produce. Lawyers, financial analysts, engineers, and other specialists define what good looks like, score the work, and turn that judgment into a model they can evaluate and deploy, so they can run the process themselves rather than waiting on an internal research team. The public launch follows use by Harvey, Rogo, Decagon, Mercor, and Nvidia, and it arrives with a $45 million Series A led by Sequoia at a $300 million valuation, with Nvidia, Bessemer Venture Partners, and Bain Capital Ventures participating. A run starts from a spec: a benchmark, a production environment, or a handful of scored examples. Trajectory handles rollouts and infrastructure, while the user can still set token budget, turn limits, reasoning effort, and the reinforcement learning algorithm. Training and deployment are each a click, the run can be inspected step by step, and the checkpoint can be compared with frontier models before it ships. The same actions are available to a coding agent. Two research previews come with the launch. Continual Grading applies a small set of human judgments across production traces, and Group Size 1 (GS1) trains on individual production interactions so a model can keep learning after it is deployed. One customer used the platform to train more than eight models for long, multi-step tasks in two weeks. Ronak Malde, Michael Elabd, and Arjun Karanam, three Stanford classmates, founded the company about a year ago. Malde, the CEO, worked on SWE-1 at Windsurf and later at Google DeepMind. Elabd also came from DeepMind, and Karanam worked on generative foundation models at Apple. The team is 18 people.