Trajectory Raises $45M to Make Post-Training Accessible to Domain Experts

Trajectory, an AI startup founded about a year ago by three Stanford classmates, has raised a $45 million Series A led by Sequoia at a $300 million valuation, according to TechCrunch. Nvidia, Bessemer Venture Partners and Bain Capital Ventures also participated in the round.

The company is building a platform for post-training, the stage in which a pretrained model is adapted through techniques such as fine-tuning (further training on a narrower dataset) and reinforcement-based optimization (adjusting behavior using a reward signal). Trajectory has made that platform publicly available. According to the company's own description, the product is aimed at users who are not machine learning specialists.

Harvey and Rogo are using the platform to post-train open-source models. Decagon, Mercor and Nvidia are listed as customers. The customer list spans legal and financial applications, customer-support automation and AI-driven recruiting data, though the company has not described the specific workloads in detail.

The three co-founders are Ronak Malde, Michael Elabd and Arjun Karanam. Malde was among the researchers Google DeepMind acqui-hired from Windsurf. Elabd previously worked at DeepMind, and Karanam previously worked on generative foundation models at Apple. The company has 18 employees.

Trajectory's website states that its researchers have brought continual learning to robotics, 3D foundation models and SWE-1. It also says its team includes product designers who have shipped products used by millions of people every day. The company has recruited researchers and product designers from OpenAI, Stripe and Netflix, according to TechCrunch.

The technical premise is that post-training has traditionally required a team fluent in training loops, data curation, evaluation harnesses and GPU scheduling. A platform that hides those layers lowers the bar for domain experts who know their data and failure modes but not the internals of optimization. Whether that holds under production demands, such as controlling for reward hacking (a model learning to game its scoring signal instead of solving the task), managing regression across checkpoints and keeping inference costs predictable, is a question for customers to answer over time. The funding announcement does not settle it.

In my view, the more useful signal in this round is the investor and customer mix. Sequoia leading alongside a chip vendor that is also a customer ties the financing to the infrastructure layer as well as to application developers. Nvidia's role as both investor and customer is one the company's disclosures describe directly, and it will be worth watching whether the platform's compute path favors any particular hardware stack.

The broader context here is a shift in who gets to shape models. For most of the past several years, fine-tuning an open-source model close to the frontier was a task for well-resourced ML teams. If a platform reliably turns that work into a configuration problem for domain specialists, the constraint moves from engineering headcount to the quality of proprietary data and evaluation criteria. That is a change with real upside for legal, financial and clinical teams that have data but no training staff. It also concentrates judgment about model behavior in whoever writes the evaluations.

Trajectory's funding gives it runway to test that thesis at a larger scale. The next measurable questions are how many post-training jobs its customers run in production, and whether the open-source models produced on its platform perform as their operators expect once they leave the sandbox.