Trajectory raises $45M for continuous learning AI

The startup Trajectory has officially announced a $45 million Series A funding round to automate the post-training of models. It promises business-driven continual learning, now accessible via a publicly open platform, in a niche already contested by several players.

A Field Already Contested by Thinking Machines Lab, Tinker, and Applied Compute

Trajectory positions itself against several players in the post-training automation niche. The startup considers direct rivals to be Thinking Machines Lab, founded by Mira Murati, former CTO of OpenAI, as well as the company behind the fine-tuning platform Tinker. It also mentions Applied Compute, valued at $3.25 billion last month, and Prime Intellect, which assists companies in creating and training their own AI agents.

$45M Funding and $300M Valuation Led by Sequoia

The company has raised $45 million in Series A funding, with a valuation of $300 million. The round is led by Sequoia, with participation from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures. The company currently employs 18 people.

A Continual Learning Platform Now Open to the Public

Trajectory focuses on continual learning: companies start with an open-source model and connect the platform to their applications to capture performance feedback. This data is then used to retrain the model to improve its skills in specialized tasks over time. The post-training platform is now available to the public. Users define what constitutes a good outcome in their organization and can make model improvements; according to Ronak Malde, specialists such as lawyers or doctors can thus intervene almost automatically. This capability could accelerate and reduce the cost of deploying specialized AI. The approach bears similarities to recursive self-improvement, but Michael Elabd emphasizes that keeping humans in the loop makes it safer, as autonomous improvement can produce unforeseen or dangerous effects. It aims to address a limitation of current models, which do not learn from their mistakes without retraining, while companies today specialize general models at the cost of internal teams or FDE consultants. Trajectory aims to automate this improvement, which Malde illustrates by the gap between a highly intelligent but still new profile and the experience gained over the years.

Harvey, Rogo, and Other Clients Fine-Tuning Open Source Models

Harvey and Rogo are already using the platform to fine-tune open-source models after their initial training. They use it to integrate learnings from previously completed tasks, such as reviewing a category of contracts or analyzing a company according to a specific method. Other clients include Decagon, Mercor, and Nvidia.

A Team from DeepMind and Apple, and a Bet on the Interface

Ronak Malde believes that Trajectory's main strength lies in its team. He is among the few researchers who joined Google DeepMind from Windsurf; Michael Elabd has also worked at DeepMind, while Arjun Karanam contributed to the development of foundational generative models at Apple. The company has also recruited talent from OpenAI, Stripe, and Netflix. According to Karanam, design plays a central role: the developed interface aims to make post-training accessible to clients without technical expertise.