Trajectory Raises $45M So Lawyers and Doctors Can Teach AI Models to Learn on the Job

Most AI models you use today have a blind spot that rarely gets talked about outside research circles: once they're trained, they don't really learn from their own mistakes. Trajectory, a young startup founded by three Stanford classmates, just raised $45 million to change that for specialists who'll never write a line of training code.

The deal

According to TechCrunch, the 18-person company closed a Series A at a $300 million valuation. Sequoia led the round, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures also taking part. The company is now opening its post-training platform to the public.

What it actually does

The idea is called continual learning. A business starts with an open source model, connects Trajectory to the app where the model is used, and the platform gathers feedback on how the model performs. That feedback is then used to retrain the model so it keeps getting better at one narrow job.

The pitch is that the experts themselves, lawyers, doctors, financial analysts, decide what a good answer looks like, instead of hiring a research team or expensive outside engineers. Co-founder Ronak Malde compared today's models to a brilliant PhD who is always starting a new job, when what people really value is someone with 30 years in the role. Harvey and Rogo already use it to tune models for legal contract review and company valuation, and Decagon, Mercor and Nvidia are listed as customers too.

The part I find interesting

This sounds a bit like recursive self-improvement, the idea of AI improving itself that's been all over the news this week. Co-founder Michael Elabd argues the difference is that humans stay in the loop, which he sees as the safer path. Trajectory isn't alone here, though: Thinking Machines Lab, Applied Compute and Prime Intellect are all chasing the same post-training market.

For anyone using AI tools at work, this is the quieter trend to watch. The next big jump might not come from a bigger model at all, but from models that keep learning on the job.