Trajectory startup raises $45M in Series A round — TechCrunch

Trajectory, a startup developing a platform for further training artificial intelligence models based on user feedback, has raised $45 million in a Series A round. The 18-person company was valued at $300 million. The round was led by Sequoia, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures also among the investors, TechCrunch reports.

A platform for post-training models

Trajectory was founded about a year ago by three Stanford classmates — Ronak Malde, Michael Elabd and Arjun Karanam. Malde was among the Windsurf researchers hired by Google DeepMind, Elabd also worked at DeepMind, while Karanam previously worked on generative foundation models at Apple.

The company offers a continuous learning approach: businesses can take an open-source model, connect it to their own applications and collect feedback on the quality of its performance. Trajectory uses this data to retrain the model so that it performs specialized tasks better over time.

According to the startup's concept, lawyers, doctors and other domain experts can independently define criteria for a model's high-quality performance and turn them into improvements to the model. Trajectory believes this could make the deployment of specialized AI faster and cheaper.

More current news is available on the UA.News Telegram channel Telegram.

Clients and competition

Harvey and Rogo already use Trajectory's platform to post-train open-source models. According to the company, this helps systems operate closer to the level of experienced lawyers or financial analysts, including by analyzing certain types of contracts or evaluating companies using defined methodologies. Other clients named include Decagon, Mercor and Nvidia.

Trajectory is also making its post-training platform publicly available. The company considers Thinking Machines Lab, which created the Tinker fine-tuning platform, as well as Applied Compute and Prime Intellect, to be competitors.

Elabd said that human involvement in training models makes this approach safer than fully autonomous self-improvement of AI systems, which could lead to unpredictable or dangerous consequences.