Trajectory raises $45M to automate post-training for AI models

Trajectory was founded about a year ago by three Stanford classmates with prior experience at Google DeepMind and Apple. The 18-person company aims to automate fine-tuning, a task that businesses often handle with internal research teams or costly consultants known as forward-deployed engineers.

Customers connect an open source model to their applications, and the platform collects feedback on how the model performs. It then uses that feedback to retrain the model so it improves at specialized work over time. Harvey and Rogo already use it to post-train models for legal and financial tasks, and Decagon, Mercor and Nvidia are also customers.

The round also included participation from Nvidia, Bessemer Venture Partners and Bain Capital Ventures. Trajectory is now making its post-training platform publicly available. The company names Thinking Machines Lab, Applied Compute and Prime Intellect as competitors, and co-founder Michael Elabd argues that keeping humans in the training loop is safer than autonomous self-improvement.

Developers building agents for specialized fields may now have a hosted option for retraining open source models from feedback instead of relying on in-house research teams for fine-tuning. Teams adopting it will need to define what good output looks like, because the platform depends on users to set that standard. Since the platform has only just become publicly available, test any performance claims on your own tasks before relying on them.