The Pulse
Trajectory Raises $45 Million as It Opens Model Post-Training Platform
Trajectory raised $45 million in a Series A led by Sequoia and is making its model post-training platform publicly available. Its customers include Harvey and Rogo, while two research previews describe methods for extending expert grading a
AI.info Team · Written with AI · fact-checked ·
In a report published November 2, TechCrunch said Trajectory had raised $45 million in a Series A at a $300 million valuation, led by Sequoia, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures also participating. The 18-person startup is pitching a way for specialists—not only machine-learning research teams—to refine models for their own work.
From general models to specialist work
Companies adapting general-purpose models for fields such as law and finance often fine-tune them for specific tasks. That work can require an internal research team or consultants known as forward-deployed engineers, TechCrunch reported. Trajectory’s product aims to let businesses start with an open-source model, connect it to their applications and use feedback on its performance to retrain it over time. The company’s website describes tools for training, evaluating and deploying model improvements, and says the platform works with any agent harness.
Trajectory co-founder Ronak Malde compared foundational models to a very smart PhD student who is always starting a new job. Malde told TechCrunch:
“But what the economy values is someone who worked on the job for 30 years,”
Malde is a co-founder of Trajectory.
Harvey and Rogo are already using it
Harvey and Rogo use Trajectory to post-train open-source models for legal and financial work. Examples include reviewing a particular type of contract or valuing a company according to a specific methodology. TechCrunch also reported that Decagon, Mercor and Nvidia are customers.
Trajectory says lawyers, doctors and other specialists can shape model improvements by defining what good performance looks like for their organization. The company’s stated aim is to reduce dependence on dedicated research teams and make specialized models quicker and less expensive to deploy. TechCrunch reported that the startup was making its post-training platform publicly available.
Two research previews address the training loop
Trajectory published work on Continual Grading, a method intended to extend expert judgments from a small set of examples to tasks that have not been graded. In the company’s research note, the team reports training a model on Harvey-LAB using generated rubrics or expert-written rubrics. The note says the results were comparable, but overlapping confidence intervals meant the researchers could not conclude that generated rubrics performed better. The experiment used 48 labeled tasks to construct the grader and another 48 to select it. It is a company-reported test on a defined benchmark, not evidence that the approach works equally well across customer settings.
A companion post describes Group Size 1, or GS1, as an approach for learning from individual production rollouts rather than requiring multiple attempts in the same task environment. Trajectory’s researchers frame it as a way to train on real-world activity that cannot easily be replayed in a simulator. Both methods are presented as research previews.
Trajectory enters a crowded post-training market
TechCrunch names Thinking Machines Lab’s Tinker platform, Applied Compute and Prime Intellect among companies pursuing automated post-training. The report says Trajectory was founded about a year ago by three Stanford classmates: Malde, Michael Elabd and Arjun Karanam. Malde and Elabd previously worked at Google DeepMind; Karanam worked on generative foundation models at Apple.
Trajectory’s bet is that businesses can turn feedback generated during ordinary product use into improvements tailored to their own work. Its platform connects model training with that feedback, while the company’s research previews offer bounded tests of two methods for grading outputs and learning from production activity.